높은 추상화  ←─────────────────────────────────→  낮은 추상화
언어 → 코드 → 추론 → 목표 → 궤적 → 어포던스 → 잠재 → 원시

VLA (Vision-Language-Action)
├── 정의 기준별 분류
│   ├── 좁은 정의 (RT-2 원조): VLM 파인튜닝 기반
│   ├── 확장 정의 (Ma et al.): V+L→A 모든 시스템
│   ├── Pure VLA (Zhong et al.): End-to-end 통합
│   └── 직접 제어 (Kawaharazuka et al.): 제어 명령 직접 생성
│
├── 아키텍처별 분류 (Liu & Shao [5])
│   ├── 단일체 (Monolithic)
│   │   ├── Single-system: RT-2, OpenVLA
│   │   └── Dual-system
│   │       ├── Cascade: GR00T N1, π0
│   │       └── Parallel: (동시 실행 후 결합)
│   └── 계층적 (Hierarchical)
│       ├── Planner-Only: SayCan, Inner Monologue
│       └── Planner+Policy: π0.5
│
├── 행동 생성 방식별 분류 (Zhong et al. [3])
│   ├── 자기회귀 (Autoregressive): RT-2, OpenVLA, Octo(AR모드)
│   ├── 디퓨전 (Diffusion): Diffusion Policy, CogACT, RDT-1B
│   │   ├── Flow Matching (변형): π0, π0-FAST
│   │   └── 이산 디퓨전 (Discrete Diffusion): 이산 토큰 공간에서의 확산
│   ├── 강화학습 기반: VLA-RL [68], RIPT-VLA [71], ConRFT [69]
│   └── 하이브리드/특수: HybridVLA [79], GR00T N1
│
├── 행동 토큰 유형별 분류 (Chen et al. [7])
│   ├── Language Tokens: SayCan, SayTap
│   ├── Code Tokens: Code-as-Policies
│   ├── Affordance Tokens: VoxPoser [57], A3VLM [58]
│   ├── Trajectory Tokens: RT-Trajectory [62], TraceVLA
│   ├── Goal Tokens: SuSIE [59], 3D-VLA
│   ├── Latent Tokens: VQ-BeT [60], LAPA [61], UniVLA [80]
│   ├── Raw Action Tokens: RT-2, OpenVLA, FAST
│   └── Reasoning Tokens: CoT-VLA [55], SC-VLA [56]
│
├── 효율화 기법별 분류 (Yu et al. [4])
│   ├── 양자화: BitVLA, SQIL
│   ├── 프루닝: SmolVLA, FLOWER, DeeR-VLA
│   ├── 증류: TinyVLA
│   ├── 토큰 최적화: FAST, VLA-Cache, VOTE
│   ├── 효율적 어텐션: KV-Efficient VLA, Long-VLA [73]
│   └── 효율적 아키텍처: SARA-RT, MoLE-VLA
│
├── 학습 패러다임별 분류 (Jin et al. [9])
│   ├── Phase 1: 인터넷 사전학습 (VLM)
│   ├── Phase 2: BC/SFT (로봇 시연)
│   └── Phase 3: RL 후처리
│       ├── 온라인 RL: PPO (RIPT-VLA [71]), GRPO (VLA-RL [68])
│       ├── 온라인 RL: ConRFT [69]
│       └── 선호 최적화: HAPO [84], GRAPE
│
└── 응용 도메인별 분류
    ├── 테이블탑 조작: 주류, 풍부한 데이터
    ├── 휴머노이드: 고DoF, 전신 제어
    ├── 자율주행: 별도 진화, 최고 안전 요구
    ├── 드론/내비게이션: 야외, 실시간
    ├── 의료/수술: 극도 정밀, 데이터 희소
    └── 산업/농업: 반복 작업, 견고성 중심
│
├── 벤치마크별 분류
│   ├── 조작: LIBERO, CALVIN, RLBench, Meta-World, VLABench
│   ├── 자율주행: Bench2Drive, nuScenes, Reason2Drive, WorldBench
│   └── 차세대: RoboArena, RoboCasa365, WorldGym

Section 1: Introduction — What Is a VLA?

1.1 Defining VLA: Three Perspectives

A Vision-Language-Action (VLA) model is a unified neural network that enables a robot to observe a scene through its cameras, understand natural-language instructions, and directly generate physical actions. However, the term "VLA" has yet to reach full consensus within the research community. Three distinct definitions currently coexist, and understanding the precise scope and meaning of each is the first gateway to navigating this field.

Narrow definition: the original meaning coined by RT-2 [11]. In 2023, Google DeepMind's RT-2 [11] paper was the first to use the term "VLA," and it carried a precise technical prescription: a model that takes a large-scale pretrained Vision-Language Model (VLM) and directly fine-tunes it for robot action prediction. RT-2 [11] took existing VLMs — PaLI-X (55B) and PaLM-E [18] (12B) — represented robot actions as text tokens, and extended the VLM's output vocabulary with action tokens. Under this definition, the essence of a VLA is "transferring internet-scale visual-language knowledge to robot actions," and the presence of a VLM backbone is a necessary condition.

Broad definition: the inclusive category of Ma et al. [1] (2024). The survey by Ma et al. [1], which offers a bird's-eye view of embodied AI as a whole, defines VLA far more expansively. For them, VLA encompasses any model that takes Vision and Language inputs and produces Actions. Under this definition, modular systems such as SayCan [14] — where an LLM handles only high-level planning while a separate policy manages low-level control — are included, as are models like RT-1 [12] that use their own architectures without a VLM backbone, and even systems like Code-as-Policy that generate code as output. This expansive definition is useful for mapping the full landscape of the field, but has been criticized for diluting the technical specificity of the term "VLA."

Pure definition: "Pure VLA" proposed by Zhong et al. [3] (2025). The most recent classification, proposed by Zhong et al. [3] to resolve the tension between the two definitions above, introduces the concept of "Pure VLA." A Pure VLA integrates perception, language understanding, and action generation within a single end-to-end sequence-modeling framework. The three key criteria are: (1) both vision and language must be used as inputs; (2) actions must be the direct output of the model (without passing through a separate low-level controller); and (3) the entire pipeline must be integrated into a single trainable model. By these criteria, SayCan [14] (modular structure) and Code-as-Policy (code output) are not VLAs, whereas models such as RT-2 [11], OpenVLA [15], and π0 [16] qualify as Pure VLAs. Zhong et al. [3] further subdivide Pure VLAs into four categories: (1) autoregressive VLAs (RT-2 [11], OpenVLA [15]), (2) diffusion VLAs (π0 [16], CogACT [23]), (3) reinforcement learning-based fine-tuning, and (4) hybrid and specialized methods. This classification scheme jointly considers both the action decoder's generation mechanism and the learning paradigm.

Beyond these three definitions, Kawaharazuka et al. [2] proposed their own boundary criterion: only systems that "take visual observations and natural-language instructions as mandatory inputs and directly generate control commands" qualify as VLAs, while "high-level policies that select from a pre-defined skill index" are explicitly excluded. This aligns with Zhong et al. [3]'s Pure VLA definition in placing SayCan [14]-style skill-selection systems outside the VLA boundary, but differs in that it uses "direct control command generation" rather than the presence of a VLM backbone as the key criterion.

This document acknowledges all four definitions, but in practice centers on Zhong et al. [3]'s Pure VLA definition. Nonetheless, modular systems (SayCan [14], Inner Monologue [22]) and hierarchical architectures (π0.5 [31], GR00T N1 [21]) are treated as important components of the broader VLA ecosystem.

1.2 Why VLA: The Essence of the Paradigm Shift

The traditional architecture of robotics has rested for fifty years on the tripartite "sense-plan-act pipeline": a perception module processes sensor data, a planning module searches for a path to the goal, and a control module generates joint commands. Each module is designed independently, and inter-module interfaces communicate through hand-crafted representations (e.g., object poses, grid maps, joint trajectories).

This modular structure offered strengths in mathematical rigor and debuggability, but it carried three fundamental bottlenecks. First, the representation bottleneck: information transmitted between modules is constrained by the expressivity of hand-crafted representations. Executing an instruction such as "put the blue plate next to the red mug in the sink" requires the perception module to detect the objects, the language module to parse the instruction, the grasping module to compute the grasp pose, and the motion planner to generate a collision-free path. Information is lost at each interface, and a failure in any one module collapses the entire pipeline. Second, the generalization bottleneck: because each module is independently tuned to specific environments or objects, adapting to a new environment or object requires re-engineering the entire pipeline. Third, the knowledge utilization bottleneck: although the internet contains billions of images and trillions of words of text, traditional pipelines have no pathway to exploit this large-scale prior knowledge.

VLA addresses all three bottlenecks simultaneously. By connecting sensor inputs to action outputs as a single differentiable function, it eliminates the representation bottleneck; and by inheriting the pretrained knowledge of internet-scale VLMs, it resolves the generalization and knowledge-utilization bottlenecks at once. The robot's process of "seeing (Vision), understanding (Language), and acting (Action)" takes place end-to-end within a single neural network.

To frame this transition with an analogy: if traditional robotics was an "international conference communicating through multiple interpreters," VLA is a "one-on-one conversation in one's native language." The errors and delays of intermediate translation disappear, and context and nuance are conveyed intact.

1.3 The Limitations of Fourteen Surveys and the Purpose of This Document

From the second half of 2024 through the first half of 2025, survey papers on VLA were published at an explosive rate. the fourteen core surveys referenced in this document are as follows:

Note: Ma et al. [1] is a living survey that has been continuously updated from its 2024 first edition through v7 (2026.02).

Each survey illuminates VLA through its own lens, yet no single survey captures the complete picture. Surveys that treat architecture in depth neglect deployment; surveys focused on efficiency omit training paradigms; surveys specialized in a particular domain miss the intersections with other domains.

This document disassembles all fourteen surveys down to the atomic level to construct a superset of their information. This is not a simple merger. Where different surveys analyze the same model from different angles, this document integrates those cross-survey perspectives to provide an understanding richer than any individual survey. For example, for π0 [16], Zhong et al. [3] contribute an architectural classification, Yu et al. [4] contribute an inference-efficiency analysis, and Jin et al. [9] contribute an RL post-processing analysis; this document weaves all three perspectives into a single unified profile.

Differentiating Contributions of This Document

If the fourteen surveys above are specialist surveys that probe specific axes, this document is a meta-survey (survey of surveys) that cross-cuts those axes. Specifically, it provides the following contributions absent from any individual survey:

1.4 Structure of This Document

This document unfolds in the following structure:

• Section 1 (this section): The definition of VLA, its significance, and the purpose of this document
• Section 2: A chronology of VLA's evolution — the developmental narrative from 2017 to 2026
• Section 3: Unified taxonomy — integrating fourteen surveys into one
• Section 4: In-depth architectural dissection — Vision Encoder, VLM Backbone, Action Decoder
• Section 5: Action tokenization — translating continuous actions into the model's language
• Section 6: Training paradigms — Behavior Cloning, pretraining strategies, RL post-processing
• Section 7: Efficiency — lightweighting and optimization for real-world deployment
• Section 8: Application domains — manipulation, humanoids, autonomous driving, drones, medical
• Section 9: Datasets, benchmarks, and simulators
• Section 10-11: Open problems, integrated insights, and conclusion

Each section synthesizes all relevant content from the fourteen surveys and explicitly provides cross-survey integrated insights.

---

Section 2: A Chronology of VLA's Evolution (2017–2026)

VLA did not emerge overnight. Three independent streams — computer vision, natural language processing, and robot learning — each carved out their own canyon over decades, until they converged at a single confluence point in the early 2020s. This section traces that confluence narrative across four phases.

Phase 0 — Convergence of Foundational Technologies (2017–2021)

The Vision Revolution: From CNN to ViT, and CLIP

After AlexNet overwhelmed traditional computer vision at the ImageNet competition in 2012, CNNs evolved rapidly. ResNet (2015) demonstrated residual learning that could deepen to 152 layers, and EfficientNet (2019) showed balanced scaling of width, depth, and resolution. Yet the true turning point came in 2020 with the Vision Transformer (ViT [28]). By dividing an image into 16×16 patches and processing them with Transformer self-attention, ViT [28] proved that the scaling laws validated in NLP applied equally to vision: more data, larger models, better performance — this simple formula fundamentally redirected vision research.

However, the development with the most direct influence on VLA was not ViT [28] but OpenAI's CLIP [27] in 2021. CLIP [27] performed contrastive learning on 400 million image-text pairs to align vision and language in the same embedding space: a photo of a cat and the text "a photo of a cat" become close together in vector space. This visual-language alignment became a core prerequisite for VLA: a robot's ability to connect the linguistic concept "red mug" with the visual object visible in the camera when given the instruction "pick up the red mug" derives directly from CLIP [27]-style pretraining. SigLIP (2023), a successor to CLIP [27], achieved more efficient training using sigmoid loss, and DINOv2 [30] (2023) learned label-free visual representations through self-supervised learning, going on to be widely adopted as the Vision Encoder in subsequent VLAs.

The Language Explosion: From GPT-3 to the Era of Large Language Models

Following the emergence of the Transformer architecture in 2017, NLP entered an era of rapid scaling. If BERT (2018) demonstrated bidirectional contextual understanding and GPT-2 (2019) demonstrated autoregressive text generation, GPT-3 [29] (2020) changed the paradigm itself with 175 billion parameters. GPT-3 [29] exhibited "emergent capabilities" — the ability to perform new tasks through few-shot prompting alone, without explicit training — and this gave robotics researchers a decisive insight: if a sufficiently large model trains on sufficiently much data, capabilities that were never explicitly programmed appear on their own.

This observation led directly to two questions. First, could the world knowledge of LLMs (common sense, physical intuition, task procedures) be leveraged for robot planning? Second, do the scaling laws of LLMs also apply to robot policies? The first question was answered in 2022 by SayCan [14] and Inner Monologue [22]; the second was answered in 2023 by RT-2 [11].

The Wall of Robot Learning: Behavior Cloning and the Data Bottleneck

At the same time, the field of robot learning was grappling with its own challenges. Behavior Cloning (BC) was the most straightforward method for learning a policy by imitating expert demonstrations, but distribution shift — the phenomenon in which errors accumulate exponentially in states not seen during training — acted as a fundamental limitation. DAgger [97] (2011) theoretically resolved this problem, but obtaining repeated expert corrections in real time in the real world was impractical.

Offline Reinforcement Learning (Offline RL) attracted attention as an alternative to BC. CQL (2020), IQL (2021), and others proposed methods for improving policies using only existing collected data. However, these also carried the limitation of conservative estimation and hyperparameter sensitivity, and presupposed large-scale datasets of tens of thousands to hundreds of thousands of episodes. The problem was that collecting robot data is incomparably slower and more expensive than in NLP or CV. While ImageNet had millions of images and GPT-3 [29]'s training data numbered hundreds of billions of tokens, the largest robot datasets at the time contained only a few thousand episodes at most.

Meanwhile, simulator ecosystems began to partially alleviate this data bottleneck. SAPIEN (2020) provided a manipulation environment based on sophisticated physics simulation, AI2-THOR (2017) provided a photorealistic home environment, and Habitat (2019) provided a large-scale navigation environment. These simulators would go on to become the core infrastructure for VLA training and evaluation.

The Key Precursor: CLIPort (2021)

The first point at which these three streams — visual-language alignment, large language models, and robot learning — crossed was CLIPort [26] (2021). Shridhar et al. combined CLIP [27]'s visual-language representations with the Transporter Network (a spatial action map for robot manipulation) to perform table-top manipulation following language instructions. CLIPort [26] was the first to empirically demonstrate the core hypothesis that "internet-scale pretrained visual-language knowledge can transfer to robot actions." Although CLIPort [26] itself was not an end-to-end VLA but a hybrid structure injecting CLIP [27] features into the Transporter, this proof of concept became the direct inspiration for RT-2 [11] and OpenVLA [15].

Phase 1 — The Birth of VLA (2022–2023)

The foundational technologies that converged in Phase 0 began to be combined explosively from 2022. This period was like crossing the activation energy of a chemical reaction: a handful of landmark models appeared in rapid succession, giving birth to an entirely new field.

Gato and SayCan: The Beginning of Two Approaches (2022)

In the first half of 2022, DeepMind simultaneously unveiled two lines of experimentation. Gato [13] (2022) was a "generalist agent" that converted text, images, robot actions, and game inputs all into tokens and trained a single 1.2B-parameter Transformer on all of them. Gato [13]'s innovation was conceptual: a proof that modalities as entirely different as these could be unified within a single sequence-modeling framework. Although its performance on each individual task fell short of specialist models, it was the first to demonstrate the possibility that "a single model can see, read, and act."

At the same time, SayCan [14] (2022) took the opposite philosophy. Rather than using an LLM (PaLM) to generate actions directly, it used it only as a high-level planner. SayCan [14]'s structure was elegant: the LLM decomposes a natural-language instruction into a sequence of step-by-step sub-tasks (e.g., "get me a Coke" → "move to the kitchen" → "open the fridge" → "pick up the Coke" → ...), pre-trained low-level policies (affordance functions) evaluate the feasibility of each sub-task, and the best executable action is selected by multiplying the LLM's plan with the affordance. SayCan [14] does not satisfy Zhong et al. [3]'s Pure VLA definition, but it stands as a key ancestor in the VLA lineage as the first system to demonstrate "connecting LLM world knowledge to a robot" in the real world.

Inner Monologue [22] (2022) advanced SayCan [14]'s idea by one step. Rather than having the LLM plan once and stop, it introduced a closed-loop structure in which the LLM receives textual environmental feedback about execution outcomes (success/failure, object detection results, human corrections) and dynamically revises its plan. This "inner monologue" is the direct ancestor of reasoning-integrated VLAs such as CoT-VLA [55].

VIMA [45]: Multimodal Prompt Following (2022)

Also emerging in this period, VIMA [45] (2022) was an encoder-decoder Transformer that understood prompts across diverse modalities — not just text, but also image goals, video demonstrations, and bounding boxes — to generate robot actions. VIMA [45] introduced the perspective that "language is not the only instruction channel," and laid the groundwork for subsequent research on multimodal instruction following.

RT-1: Proof of Large-Scale Real-World Learning (2022)

If SayCan [14] was a strategy for borrowing the LLM's "wisdom," RT-1 [12] (Robotics Transformer 1, 2022) took a head-on approach. Google spent 17 months collecting 130,000 real-world demonstration episodes with 13 robots, and trained them on a approximately 35M-parameter Transformer model. RT-1 [12]'s architecture was comparatively modest (EfficientNet + TokenLearner + Transformer decoder). Yet what RT-1 [12] proved was not its architecture but scale: training on sufficiently diverse and large-scale real-world data allows a single model to perform more than 700 different tasks.

RT-1 [12] also revealed an important failure: generalization to objects or environments not seen in the training data was extremely limited. This limitation led directly to the question "could we inject knowledge from internet-scale pretraining?", and the answer was RT-2 [11].

RT-2 and PaLM-E: The Official Birth of VLA (2023)

2023 was the year VLA got its name. RT-2 [11] (2023) took the existing VLMs PaLI-X (55B) and PaLM-E [18] (12B), encoded robot actions as text tokens (discretizing each action dimension into 256 bins and representing them as integer strings), and jointly fine-tuned them by mixing a small amount of robot data into the VLM's existing training data. The results were dramatic: even though RT-2 [11] was trained on the same robot data as RT-1 [12], it generalized to objects not seen during training ("put the dinosaur in the correct bin") and to abstract concepts ("pick up the object that is different from the others"). The internet-scale knowledge of the VLM had transferred to robot actions.

PaLM-E [18] (2023), published nearly simultaneously, was a massive multimodal model with 562B parameters that integrated visual tokens, language tokens, and robot-state tokens into a single input sequence. PaLM-E [18] focused more on multimodal understanding and planning than on direct action generation, but it was the most extreme push of the scaling hypothesis that "a single giant Transformer can digest vision, language, and robot actions all at once."

The message from RT-2 [11] and PaLM-E [18] was clear: a VLM is not merely an image captioner — with appropriate fine-tuning, it can become a policy that generates actions in the physical world. This is the core proposition of the VLA paradigm, and 2023 — the year this proposition was empirically demonstrated — is the founding year of the VLA field.

Diffusion Policy: A New Grammar for Action Generation (2023)

Almost simultaneously with RT-2 [11]'s proposal of an approach that autoregressively generates actions as discrete tokens, Diffusion Policy [17] (Chi et al., 2023) opened an entirely different action-generation paradigm. It applied the DDPM (Denoising Diffusion Probabilistic Model) — which had revolutionized image generation through DALL-E and Stable Diffusion — to robot action generation.

The core insight of Diffusion Policy [17] lay in the multimodality of robot actions. For the instruction "pick up the cup," there is no single correct action — the robot might grasp from the right, from above, or by the handle. Training with a conventional mean squared error (MSE) loss would learn the average of this multimodal distribution, generating meaningless actions belonging to none of the modes. Diffusion Policy [17] naturally represented the multimodal distribution by starting from Gaussian noise and generating an action sequence (action chunk) through iterative denoising. Furthermore, its natural compatibility with action chunking -- generating entire action sequences at once -- was a key contribution in securing temporal consistency.

This approach had a fundamental influence on the design of the Action Decoder in subsequent VLAs. The Flow Matching in π0 [16], the DiT-based decoder in CogACT [23], and the diffusion Transformer in RDT-1B [24] all developed on the grammar that Diffusion Policy [17] opened.

Open X-Embodiment: A Turning Point in Data Integration (2023)

In late 2023, the Open X-Embodiment [19] (OXE [19]) project, led by Google DeepMind, fundamentally transformed the data infrastructure of the VLA field. This dataset, which integrated more than one million episodes collected from 22 robotic platforms into a standardized format (RLDS), shattered the paradigm in which individual research labs collected data only from their own robots. OXE [19]'s key finding was that policies trained on "cross-embodiment data" generalized better than policies trained on single-robot data. Data from different robots acted as "diversity" rather than "noise," preventing overfitting.

OXE [19] became the training data foundation for nearly all subsequent large VLAs. Octo [25], OpenVLA [15], and π0 [16] all used OXE [19] as their core training data, and the very existence of OXE [19] made the research direction of "generalist robot policies" possible.

Phase 2 — Diversification and Explosive Growth (2024)

2024 was the Cambrian explosion of the VLA field. Building on the two paradigms proven by RT-2 [11] and Diffusion Policy [17] — autoregressive token generation and diffusion-based action generation — dozens of new models emerged, rapidly reshaping the landscape of the entire field.

The Emergence of Generalist Policies: Octo and OpenVLA

Octo [25] (2024) was the first generalist cross-platform policy to fully leverage the OXE [19] dataset. A relatively small model at 93M parameters, it featured a Transformer-based architecture with a flexible design supporting both autoregressive and diffusion action heads. Octo [25]'s core contribution was less in architectural innovation than in establishing the training recipe of "cross-robot pretraining → target robot fine-tuning." By demonstrating that a new platform could be adapted to with only a small amount of target robot data, it empirically validated the transfer learning paradigm for VLA.

OpenVLA [15] (2024) was a turning point in the democratization of VLA. Whereas RT-2 [11] was a proprietary model with 55B parameters, OpenVLA [15] was a fully open-source model at 7B parameters — fine-tuned on OXE [19] data using a Llama 2-based VLM — that provided a VLA anyone could reproduce. OpenVLA [15] showed an absolute success rate 16.5% higher than RT-2-X while reducing model size to approximately one-seventh. This result provided the important implication that "VLA does not necessarily require tens or hundreds of billions of parameters," and laid the groundwork for subsequent research into small VLAs.

New Architectural Paradigms: π0 and CogACT

The most influential architectural innovation of 2024 came from π0 [16] (Physical Intelligence, 2024). π0 [16] proposed a new architecture that mediates between the autoregressive approach of the RT-2 [11] family and the diffusion approach of Diffusion Policy [17]: a VLM backbone (PaliGemma-based, ~3B parameters) handles vision and language understanding, and on top of it a separate Action Expert (~0.3B parameters) generates continuous actions via Flow Matching, for a total of approximately 3.3B parameters. Unlike DDPM, which iterates stochastic reverse processes, Flow Matching directly learns a deterministic ODE path (velocity field) from noise to data, enabling faster and more stable action generation.

π0 [16]'s true impact came not only from its architecture but from its performance. By decisively outperforming all prior VLAs on long-horizon, complex manipulation tasks such as folding shirts and cleaning a table, the combination of "VLM backbone + diffusion/flow action decoder" became the new reference point for VLA architecture.

CogACT [23] (2024) used a DiT (Diffusion Transformer) as the Action Decoder and introduced a technique for adaptively ensembling action candidates generated from multiple denoising paths. RDT-1B [24] (2024) used a 1B-parameter DiT as a diffusion policy, demonstrating that a Scalable Diffusion Transformer is also effective for action generation in VLA. GR-2 (2024) leveraged large-scale web video as pretraining data, proposing a strategy that circumvents the scarcity of robot data through video pretraining.

FAST Tokenization: A Breakthrough in Action Sequence Compression

Another key innovation, developed in late 2024 and released in early 2025, was FAST [20] (Fast Action Tokenization). The method previously used in VLAs to convert continuous actions into tokens (bin discretization) was extremely inefficient — representing an action chunk (16 timesteps) for a 7-DoF robot required 112 tokens. FAST extracts the frequency components of an action sequence using the Discrete Cosine Transform (DCT) and compresses repeated patterns using Byte Pair Encoding (BPE), representing the same action sequence in up to approximately one-thirteenth the number of tokens (up to ~13x compression). This compression directly improved the inference speed of autoregressive VLAs and would become the core foundational technology of π0-FAST [20].

Combining 3D Understanding and World Models

3D-VLA [37] (2024) was a pioneering attempt to integrate 3D spatial understanding into VLA. It sought to overcome the fundamental limitations of 2D image-based VLAs — the absence of depth perception and the inability to reason about occluded objects — through a generative 3D world model. 3D-VLA [37] proposed a structure that predicts the future state of the 3D scene before executing an action, and feeds this prediction back into action generation. This approach went on to develop into SpatialVLA [39], PointVLA [77], and others, opening a research direction that adds "imagination" to the "perception-action" loop.

Phase 3 — Efficiency and Deployment Readiness (2025–2026)

If the Cambrian explosion of 2024 explored "what is possible," research from 2025 onward shifted its center of gravity toward "how to make it practical." This transition proceeded simultaneously along three axes: hierarchization of architectures, extreme efficiency, and the internalization of safety and reliability.

The Rise of Hierarchical Architectures

GR00T N1 [21] (NVIDIA, 2025) proposed an architecture inspired by the Dual Process Theory of human cognitive science. System 2 (VLM-based, 10 Hz) handles high-level understanding and planning, while System 1 (diffusion-based, 120 Hz) generates reflexive low-level actions. This separation arose from a practical insight: VLM reasoning is slow but rich, while diffusion generation is fast but simple. Combining the two systems achieves both the depth of high-level understanding and the responsiveness of low-level control.

π0.5 [31] (Physical Intelligence, 2025) adopted a different form of hierarchization. A high-level VLM generates a sequence of natural-language sub-tasks (e.g., "grab the cup" → "move over the sink" → "set the cup down"), and a low-level π0 [16] executes each sub-task. This approach opened a path to handling long-horizon tasks exceeding 30 minutes (such as cleaning an entire kitchen) with VLA.

The Extreme Pursuit of Efficiency

The most direct barrier to real-world VLA deployment was computational cost. Inference with a 7B-parameter model requires 16–24 GB of VRAM and imposes latency of hundreds of milliseconds — both fatal for robot control. 2025 was a year in which solutions to this problem were proposed at an explosive rate.

SmolVLA [32] (2025, 450M) answered the question "can VLA work without a large VLM?" With 450M parameters, it enables single-GPU training while achieving performance close to OpenVLA [15] (7B) on simple tasks. BitVLA [33] (2025) went further, applying 1.58-bit ternary quantization to VLA to dramatically reduce memory usage. TinyVLA [34] (proposed in 2024, a pioneer of the efficiency trend) proposed distillation techniques for shrinking the VLM backbone while maintaining performance; EdgeVLA (2025) proposed optimizations for edge-device deployment; and VLA-Cache (2025) proposed eliminating redundant computation by caching visual tokens. DeeR-VLA [35] (2025) adopted dynamic inference that activates layers of different depth depending on the difficulty of the input, while MoLe-VLA (2025) pursued parameter efficiency through a Mixture-of-Experts structure.

The common message from these efficiency research efforts is clear: the core value of VLA lies in the knowledge of large VLMs, but delivering that knowledge does not necessarily require a large model. Through distillation, quantization, caching, and dynamic inference, the knowledge of large models can be compressed into small models and shaped into a form suitable for real-time robot control.

The Emergence of RL Post-Processing

To overcome the fundamental limitation of VLAs trained solely with Behavior Cloning — that demonstration quality becomes the performance ceiling and behaviors not in the demonstrations cannot be discovered — 2025 saw a full-scale emergence of research on post-training pretrained VLAs with reinforcement learning. VLA-RL [68] (2025) proposed GRPO (Group Relative Policy Optimization); ConRFT [69] (2025) proposed online RL; SimpleVLA-RL [70] (2025) proposed a simplified REINFORCE-based RL; and RIPT-VLA [71] (2025) proposed iterative RL-based refinement.

These works all commonly follow the recipe of "pretrain with BC to obtain a reasonable initial policy, then explore and improve with RL to achieve performance that surpasses demonstrations." This is structurally identical to the development trajectory in the LLM field from GPT-3 [29] (pretraining) → InstructGPT (RLHF post-training), showing that the VLA field is rapidly tracking LLM's maturation path.

Integration of Reasoning and Internalization of Safety

CoT-VLA [55] (2025) introduced Chain-of-Thought reasoning from LLMs into VLA. Rather than generating actions immediately, it first generates a visual reasoning process (key region masks, task decomposition text) and then generates actions conditioned on this. This is an attempt to overcome VLA's "reactive" limitation and represents the latest evolution in the lineage of "thinking robots" that began with Inner Monologue [22] (2022).

SafeVLA [75] (2025) is the first model to internalize safety constraints into VLA's training process. Whereas existing VLAs handled safety as a post-hoc filter, SafeVLA [75] includes safety-violation prediction as part of the training objective itself, suppressing dangerous actions at the generation stage. This approach is the first systematic response to the warning that "hallucination in VLA means physical accidents."

Humanoid-VLA (2025) extended the application range of VLA from table-top manipulation to whole-body humanoid control. Controlling the full-body motion of a humanoid with dozens of degrees of freedom through VLA is a challenge of an entirely different order from conventional robot arms (6–7 DoF) in terms of action space dimensionality.

Autonomous Driving VLA: A New Application Frontier

The most dramatic demonstration that VLA applications are not limited to robot manipulation came from the autonomous driving (AD) domain. EMMA [46] (Waymo, 2024) applied the Gemini VLM to autonomous driving, presenting the possibility of an autonomous-driving VLA that end-to-end generates driving paths from sensor inputs. ORION [47] (2025) integrated visual reasoning with driving action generation; AutoVLA [48] (2025) proposed a VLA architecture specialized for autonomous driving; and DriveMoE [49] (2025) proposed an efficient structure that activates specialized modules per driving scenario using Mixture-of-Experts.

Autonomous-driving VLAs share their technical DNA with robot manipulation VLAs (VLM backbone, action tokenization, end-to-end learning), but domain characteristics differ substantially: higher speeds, stricter safety requirements, and greater environmental variability. The cross-pollination between these two domains is accelerating the maturation of VLA technology.

Explosive Quantitative Growth of VLA Research: Evidence from ICLR 2026

The most dramatic evidence of the growth rate of the VLA field comes from the submission statistics of ICLR 2026 (Reuss, 2026). At ICLR 2024, there was only a single VLA-related submission (rejected); at ICLR 2025, there were 9; but at ICLR 2026, 164 papers were submitted, representing an 18-fold year-over-year explosion. This figure signals that VLA is no longer a niche research topic but has established itself as a mainstream research direction within the machine learning community. Some analyses project over 1,000 submissions at ICLR 2027.

The key trends observed across these 164 papers include:

• Discrete Diffusion VLA: Four concurrent papers replacing the slow sequential generation of autoregression with parallel diffusion
• Embodied Chain-of-Thought (ECoT [92]): Integrating spatially-grounded reasoning with action generation
• Cross-Action-Space Learning: Transfer across heterogeneous embodiments via X-VLA [53], XR-1, HiMoE-VLA [54], and others
• Self-Improving RL: Residual RL achieving 99% on LIBERO, accelerating benchmark saturation
• New Benchmarks: RoboArena (real-sim transfer), RoboCasa365 (365 tasks / 2000+ kitchen scenes), WorldGym (world-model-based evaluation)

A particularly noteworthy finding comes from VLM4VLA (ICLR 2026), which demonstrated that standard VLM benchmark performance has no correlation with downstream VLA performance. This implies that the choice of VLM backbone for VLA should be guided not by general VLM benchmark rankings but by criteria specialized for robotic tasks.

Platform-Scale Orchestration

Another hallmark of 2025–2026 is that VLA is evolving beyond the level of individual models to the platform level. Gemini Robotics (Google DeepMind, 2025) placed Gemini 2.0 at the center of robot control, proposing a structure in which a single VLM orchestrator coordinates diverse robotic platforms and tasks. NVIDIA's GR00T ecosystem is constructing a full-stack pipeline spanning Cosmos (simulation) → GR00T N1 [21] (VLA policy) → Jetson (edge inference).

This platform strategy signals that VLA is no longer a purely research topic but is transitioning into an industrial product. Physical Intelligence's π series (π0 [16] → π0.5 [31] → π0-FAST [20]) positions itself as the "Android for robots," and Figure AI's embedding of its Helix VLA in its own humanoid to pursue a full-stack robotics company is in the same vein.

The Frontier-vs-Open-Weight Gap

The most salient dividing line as of 2025-2026 is the real-world generalization gap between closed frontier models and open-weight research models. Closed models such as Gemini Robotics and π0.5 [31] are demonstrating zero-shot real-world generalization, while open-source research VLAs, despite approaching parity on simulation benchmarks, have not narrowed the gap in real-world settings. Three factors are cited as root causes of this divide: (1) differences in training data quality and diversity, (2) a ceiling effect in simulation benchmarks that masks actual progress, and (3) disparities in research infrastructure scale. Based on this analysis, Reuss (2026) draws a critical conclusion: the current academic focus on pushing numbers on saturated benchmarks such as LIBERO and SimplerEnv risks masking the real-world deployment gap rather than closing it. The true breakthrough paths, he argues, lie in (1) data curation and quality control, (2) strengthening in-context learning capabilities, and (3) establishing real-world evaluation protocols. He further notes that data curation and in-context learning were the most underrepresented research directions at ICLR 2026 — which he frames as precisely where the greatest opportunity lies.

The Key Transition: From "Proof of Concept" to "Deployment Readiness"

VLAs of 2022–2023 were in the stage of proving "this is possible." The essence of that era was RT-2 [11] showing that VLM knowledge can transfer to robot actions, and Diffusion Policy [17] opening the possibility of multimodal action generation. 2024 was an era of explosive exploration of "how many different ways is it possible?" Dozens of models experimented with different architectures, training strategies, and application domains.

And in 2025–2026, the VLA field is converging on the question "how do we make it work in the real world?" The forces driving this transition are multilayered: lightweighting technology lowers the physical barriers to edge deployment; RL post-processing breaks the performance ceiling of BC; internalization of safety constraints structurally resolves reliability issues; and hierarchical architectures address the challenge of long-horizon, complex tasks. The simultaneous advancement along these four axes is the driving force converting VLA from a laboratory demo to a real-world product.

The most important pattern observed in this chronology is the acceleration of cross-pollination between technologies. Just as CLIP [27]'s visual-language alignment gave birth to CLIPort [26], GPT-3 [29]'s few-shot learning gave birth to SayCan [14], and image diffusion models gave birth to Diffusion Policy [17], every core innovation in VLA has been the result of adapting a breakthrough from an adjacent field to robotics. This pattern will continue: LLM reasoning techniques (CoT, MCTS) are already being transplanted into VLA, advances in video generation models will accelerate world model-based VLA, and advances in multi-agent LLM systems will lead to multi-robot VLA collaboration.

The history of VLA is still in its opening chapter. Yet the density of this opening chapter — reaching from proof of concept to industrial deployment readiness in just four years — gives us a sense of the pace at which this field will unfold going forward.

---

Motivation Chain: The Causal Lineage of Key VLA Models

---

Discriminative Features: Easily Confused Model Pairs

---

Intuitive One-Liners

• RT-2 [11]: "Just as Google Translate converts Korean to English, RT-2 makes a VLM translate images into robot actions."
• OpenVLA [15]: "Take the core idea of RT-2, shrink it 7x, and release it as open source for everyone."
• π0 [16]: "A VLM understands the situation, and a diffusion model paints smooth, precise motions based on that understanding."
• SayCan [14]: "ChatGPT gives the recipe; the robot chef actually cooks — knowing what to do and being able to do it are separated."
• Diffusion Policy [17]: "Just as Stable Diffusion paints images from noise, this paints robot motions from noise."
• Octo [25]: "A 'universal driver's license' pretrained on diverse robot data — just a small fine-tune adapts it to a new robot."
• OXE [19]: "Just as ImageNet transformed CV, OXE is robotics' ImageNet — 1M+ episodes from 22 robot types, unified."
• FAST: "Compress robot actions the way JPEG compresses images — frequency-domain encoding that cuts token count by up to 13x."
• GR00T N1 [21]: "A slow but deep-thinking cerebrum (VLM, 10 Hz) and a fast-reacting cerebellum (diffusion, 120 Hz) working in tandem."
• CoT-VLA [55]: "A robot that reasons 'why should I do this?' before acting — the evolution from reflex to deliberation."

---

Self-Check Questions: Sections 1–2

Q1: Among the three (+one) definitions of VLA, under which definitions is SayCan classified as a VLA, and under which is it excluded?

Answer: Under Ma et al.'s broad definition, SayCan is included (any system that takes vision + language and produces actions). However, under Zhong et al.'s Pure VLA definition, it is excluded (its modular structure fails the end-to-end integration criterion). Under Kawaharazuka et al.'s definition, it is also excluded (as a high-level policy selecting from pre-defined skill indices, it does not satisfy the "direct control command generation" criterion). Under RT-2's narrow definition, it is likewise excluded (the VLM backbone is not directly used for action generation).

Q2: Why did RT-2 exhibit better generalization than RT-1 despite being trained on the same robot data?

Answer: RT-2 inherited the visual-language knowledge of a VLM (PaLI-X, 55B) that had been pretrained at internet scale. Because the VLM already understood abstract concepts such as "dinosaur" and "the object that is different from the others," it could generalize to objects and concepts never encountered in the robot training data by leveraging the VLM's prior knowledge. This is the core proposition of the VLA paradigm: "transfer of internet-scale knowledge to robot actions."

Q3: Explain, from the perspective of "multimodality," why Diffusion Policy is superior to conventional MSE-based Behavior Cloning.

Answer: For the instruction "pick up the cup," there are multiple valid actions (grasping from the right, from above, by the handle, etc.). Training with an MSE loss learns the mean of this multimodal distribution, producing meaningless intermediate actions that belong to none of the modes (mode averaging). Diffusion Policy generates actions through iterative denoising from noise, naturally sampling from each mode of the multimodal distribution. Additionally, action chunking (generating an entire future action sequence at once) ensures temporal coherence.

---

Open Research Questions: Sections 1–2

Section 3: A Unified Taxonomy — Integrating Fourteen Surveys into One

"Like blind men touching different parts of an elephant, each survey was describing a different region of the vast animal called VLA. This chapter joins their hands to reconstruct the elephant's entire body."

Between late 2024 and early 2026, more than fourteen survey papers on VLA were published in rapid succession. Each survey proposed its own taxonomy, and it was not uncommon for the same model to be placed in different categories across papers. Some classify RT-2 [11] as a "monolithic VLA," while others classify it under "autoregressive action generation." The goal of this chapter is to reinterpret these taxonomies not as competing alternatives but as complementary perspectives, and to construct a meta-taxonomy capable of locating every VLA model within a single coordinate system.

---

3.1 Architectural Perspective — Based on Liu/Shao (2025)

Liu and Shao's survey focuses on the structural form of VLAs. Their central question is: "In what relationship are the VLM and the action generator connected?"

3.1.1 Monolithic Architecture

The entire system is composed of a single end-to-end model, with boundaries between internal modules emerging naturally during training.

Single-system: Observation-to-action generation occurs in a single unified forward pass. Given input images and language instructions, the network directly produces robot actions as output. Representatively, RT-2 [11] interprets the output tokens of the PaLI-X VLM directly as action tokens, and OpenVLA [15] repurposes the language model head of the Prismatic VLM for action prediction. NORA likewise integrates vision, language, and action within a single transformer.

Dual-system: Two distinct modules — a VLM backbone (System 2) and an action expert (System 1) — exist and cooperate within a single model. This distinction is inspired by Daniel Kahneman's dual-process theory. The VLM handles slow, deliberative thinking (System 2) for scene understanding and reasoning, while the action expert handles fast, reactive action generation (System 1).

Dual-system models are further divided by their information flow pattern:

• Cascade-based: The VLM runs first to produce a feature representation, which is then passed sequentially to the action expert. In CogACT [23], the VLM extracts visual-language features, after which a separate diffusion-based action generator receives them to output action sequences. In GR00T N1 [21], the Eagle-2 VLM produces context embeddings that are passed to a DiT (Diffusion Transformer) action head. Fast-in-Slow also follows a sequential structure of slow VLM processing followed by fast action generation.

• Parallel-based: VLM tokens and action tokens are processed simultaneously through a shared attention mechanism. In π0 [16], the tokens from the PaliGemma VLM and the flow-matching tokens from the action expert undergo cross-attention within shared transformer blocks. π0.5 [31] extends this so that high-level planning and low-level action are processed within the same attention space. GraspVLA [81] adopts a structure where grasp-specialized tokens are processed in parallel with VLM tokens.

3.1.2 Hierarchical Architecture

Planning and execution are explicitly separated. The higher level decides "what to do," while the lower level decides "how to move."

• Planner-Only: The VLM generates only a plan; action execution is delegated to a separate low-level controller (MPC, PID, etc.). SayCan [14] is the seminal example, followed by COME-Robot, Inner Monologue [22], and others in this lineage.
• Planner + Policy: A VLM planner generates an intermediate representation, which a learned policy then translates into low-level actions. VoxPoser [57], RT-H, and RoboPoint fall into this category.

These can be further classified by the form of intermediate representation:

• Keypoint (K): Goal specification via keypoint coordinates (RoboPoint, Rekep)
• Subtask (S): Linguistic description of subtasks (SayCan [14], ProgPrompt)
• Program (P): Generation of executable code/programs (Code-as-Policies [36], VoxPoser [57])

In addition, Liu & Shao [5] identify Affordance (A) as a separate auxiliary representation type. A3VLM [58], CoA-VLA, and similar models combine affordance maps with other representations (K, S, P) to explicitly specify graspable regions.

---

3.2 Action Generation Perspective — Based on Zhong et al. (2025)

Zhong et al. [3]'s survey shifts the lens of classification from the overall architectural form to how actions are generated. This perspective is practically important because, even with the same VLM backbone, different action generation methods can yield vastly different performance characteristics.

Autoregressive Method

Continuous actions are converted into discrete tokens, then sequentially generated using the same next-token prediction as language models. RT-2 [11] pioneered this approach with 256-bin quantization, followed by OpenVLA [15], Octo [25] (which supports both AR and diffusion modes), and RT-2-X. FAST (Fast Action Tokenization) [20] applies DCT + BPE to preserve the autoregressive framework while reducing information loss and accelerating pretraining by 5x.

• Advantages: LLM infrastructure (KV cache, quantization, speculative decoding, etc.) can be reused directly. Implementation is simple.
• Disadvantages: Quantization errors accumulate, and multimodal action distributions (e.g., situations where an object can be rotated either left or right) are hard to represent. Sequential token-by-token generation is slow for high-frequency control.

Diffusion Method

Probabilistic generative models such as DDPM, DDIM, Flow Matching, and VAE are used to sample from the action distribution. Diffusion Policy [17] was the first to demonstrate the potential of DDPM-based action generation. CogACT [23] reduced denoising steps using DDIM, and π0 [16] achieved faster and more stable generation via Flow Matching with ODE-based trajectories.

• Advantages: Naturally represents multimodal distributions. Generates smooth trajectories. Operates directly in continuous space with no quantization error.
• Disadvantages: Inference is slow due to multiple denoising steps (DDPM: 50–100 steps, DDIM: 10–20 steps, Flow Matching: 5–10 steps). Technical effort is required to ensure training stability.

Discrete Diffusion Method

Discrete Diffusion VLAs: A novel paradigm proposed simultaneously by four independent studies at ICLR 2026. Unlike conventional diffusion methods that operate in continuous space, discrete diffusion is applied directly to tokenized action sequences, generating discrete tokens in parallel without the sequential generation of autoregressive methods. dVLA [65], DIVA, and UNIFIED DIFFUSION VLA belong to this category, reporting 95–98% success rates on LIBERO. This approach attempts to combine the interpretability of autoregressive methods with the multimodality of diffusion, and has emerged as one of the most active research directions in 2026.

Reinforcement Learning (RL) Method

The policy is directly optimized using a reward signal. Pure RL-based VLAs are rare, but the pattern of fine-tuning BC-pretrained VLAs with RL is rapidly emerging. Representative examples include GRPO (Group Relative Policy Optimization), RLVF (Reinforcement Learning from Visual Feedback), and π^*_0.6 [157].

Hybrid Method

Combines the strengths of autoregressive and diffusion methods. HybridVLA [79] proposes a dual-decoding structure in which high-level semantic tokens are generated autoregressively and low-level continuous actions are generated via diffusion. UniVLA [80] combines latent world model representations and action generation within a single framework.

Specialized Method

Architectures designed for the requirements of specific domains: 3D point cloud input (PointVLA [77], 3D-VLA [37]), tactile sensor integration (ForceVLA [78], TactileVLA), cross-embodiment learning (Octo [25], CrossFormer [50]), and autonomous driving specialization (EMMA [46], DriveVLM [91]).

Cross-Action-Space Learning: Research on bridging the action-space differences among heterogeneous robot morphologies (arms, humanoids, mobile robots, etc.) is also gaining momentum. X-VLA [53] conditions on heterogeneous embodiments via soft prompting tokens, XR-1 introduces Unified Vision-Motion Codes (UVMC), and HiMoE-VLA [54] assigns action-space-specific experts through a hierarchical Mixture-of-Experts architecture.

Autonomous Driving VLA Taxonomy (Hu et al., 2025): Domain-specific VLA taxonomies are also advancing, particularly for autonomous driving. Hu et al. distinguish two AD-VLA paradigms: (1) End-to-End VLA — integrating perception, reasoning, and planning within a single model (with sub-categories of textual action vs. numerical action), and (2) Dual-System VLA — separating slow deliberation (VLM) from fast safe execution (planner) (with sub-categories of explicit guidance vs. implicit representation transfer). While structurally similar to Liu & Shao [5]'s monolithic/hierarchical classification, this taxonomy is differentiated by placing the safety–real-time trade-off as its central axis.

In particular, Hu et al. [41] provide a more granular AD-VLA taxonomy than Jiang et al. [10]. They sub-divide End-to-End VLA into textual action and numerical action categories, and partition Dual-System VLA into explicit guidance and implicit representation transfer, reflecting the unique safety-efficiency trade-offs of the autonomous driving domain. They also propose WorldBench, a unified evaluation platform that enables open-loop and closed-loop assessment within a single framework.

---

3.3 Anatomical Perspective — Based on Xu et al. (2025)

Xu et al. [8]'s survey dissects VLA through a biological analogy. Their framework consists of three organs:

• Perception — the robot's sensory organs: visual encoders (SigLIP, DINOv2 [30], CLIP [27]), proprioception encoders (proprioception MLP), and tactile/depth/force sensors convert information from the external world into internal representations.
• Brain — the central nervous system: the VLM backbone integrates perceptual information and language instructions to perform "understanding" and "planning." The brain has evolved from pure transformers to VLMs and further to VLMs capable of CoT reasoning.
• Action — the motor nervous system: translates the brain's intentions into physical movements. Discrete token heads, diffusion heads, and flow matching heads fall into this category.

The strength of this perspective is its intuitive clarity. Any VLA model can be described by three questions: "What eyes does it have?", "What brain does it have?", and "What hands does it have?" For example, π0 [16] is a model with "eyes of SigLIP, a brain of PaliGemma, and hands of Flow Matching."

Xu et al. [8]'s contribution extends beyond modular anatomy. Their central contribution is a systematic Five Challenges framework for VLAs: (1) Representation -- how to unify visual, linguistic, and action representations, (2) Execution -- how to achieve stable and precise action generation, (3) Generalization -- how to guarantee transfer to novel environments, objects, and tasks, (4) Safety -- how to prevent dangerous behaviors in the physical world, and (5) Dataset & Evaluation -- how to design fair and reproducible benchmarking. These five challenges provide the systematic foundation for the discussions of generalization, efficiency, safety, and benchmarking in Sections 7--10.

---

3.4 Functional Perspective — Based on Kawaharazuka et al. (2025)

Reflecting the practical tradition of the Japanese robotics community, Kawaharazuka et al. [2] classify VLAs not by their components but by the functions they perform:

• Low-level Perception: Raw sensory processing such as object detection, depth estimation, and pose estimation
• High-level Perception: Semantic processing such as scene understanding, relational reasoning, and affordance recognition
• High-level Planning: Task decomposition, subgoal setting, and abstract action sequence generation
• Low-level Planning: Concrete trajectory generation, motor command planning, and collision avoidance
• Data Augmentation: Simulation data generation, data augmentation via video prediction, language re-labeling, etc.

The distinctive value of this perspective is that it naturally captures the fact that a single model can simultaneously perform multiple functions. RT-2 [11] handles both high-level perception and low-level planning within a single model, while SayCan [14] specializes in high-level planning and delegates low-level execution to a separate policy.

---

3.5 Post-processing Perspective — Based on Jin et al. (2025)

Jin et al. [9]'s survey focuses on the process of adapting a pretrained VLM for robotic action generation. Their question is: "How do we transfer the knowledge a VLM already possesses to a robot?"

• Environment Perception Enhancement: Strengthening the VLM's visual understanding for the robot's environment. This includes integrating depth information, handling multiple views, and adding temporal context. SpatialVLA [39]'s depth integration and HPT [96]'s multi-camera processing are representative examples.
• Embodiment Awareness Improvement: Injecting the physical characteristics of the robot itself (joint structure, action space, dynamics) into the VLM. This includes proprioception tokenization, cross-embodiment learning, and robot-specific adapters.
• Task Understanding Deepening: Elevating the understanding of language instructions from simple semantic matching to inferential comprehension. This includes CoT reasoning (ECoT [92], CoT-VLA [55]), subgoal decomposition, and visual reasoning.
• Multi-component Integration: Methodologies that integrate the above three dimensions into a single framework. Strategies such as multitask learning, modular architectures, and continual learning are employed.

---

3.6 Meta-classification — Classifying the Taxonomies

Here we arrive at the central insight of this chapter. The five taxonomies above are not in competition with each other. They are aerial photographs of the same landscape taken from different altitudes. Just as an architect draws a building using structural, plumbing, electrical, and aerial plans, each survey captures a different cross-section of the complex system that is VLA.

Correspondences Among Classification Dimensions

The following table shows how each taxonomy describes the same models:

Complementary Classification Axes

The pattern revealed by this table is clear. Each taxonomy describes complementary axes:

Accordingly, every VLA model can be represented as a point in this 5-dimensional coordinate space. For example, the coordinates of π0 [16] are:

π0 [16] = (Parallel Dual-system, Flow Matching, PaliGemma+SigLIP+FlowHead, High-level Perception+Low-level Planning, Multi-component Integration)

The practical value of this meta-classification is threefold. First, when a new VLA model appears, it can be immediately positioned on the five axes. Second, unexplored combinations (e.g., "hierarchical structure + Flow Matching + tactile enhancement") can be systematically identified. Third, conclusions from different surveys can be integrated and interpreted without contradiction.

Meta-patterns Within the Taxonomies Themselves

Stepping back further to observe the taxonomies themselves, an interesting meta-pattern emerges:

• Structure-centric taxonomies (Liu/Shao [5], Xu) answer: "How do you build this model?" — the engineer's perspective
• Process-centric taxonomies (Zhong, Jin) answer: "How do you train this model?" — the researcher's perspective
• Function-centric taxonomy (Kawaharazuka) answers: "What can this model do?" — the user's perspective

The convergence of these three perspectives is itself an indicator of the maturation of VLA research. As a field matures, the methods of building, training, and deploying models develop independently while forming a mutually coherent system.

---

---

Section 4: In-Depth Anatomical Dissection of VLA Architecture

"VLA is an organism composed of three modules. The eyes see the world, the brain understands it, and the hands act. In this chapter, each organ is placed on the dissection table."

---

4.1 Perception Module — The Robot's Eyes

The first module of a VLA is the perception module, which transforms raw sensory data into meaningful internal representations. Just as the human visual cortex converts photons from the retina into concepts like "red cup," "on the table," and "tilted," the perception module transforms pixel arrays into token sequences interpretable by the robot.

4.1.1 Language-supervised Encoders

CLIP [27] (Contrastive Language-Image Pretraining) and SigLIP (Sigmoid Loss for Language-Image Pretraining) are encoders trained via contrastive learning on image-text pairs. Because they were trained on hundreds of millions of image-caption pairs, their visual representations are inherently semantic. "Red cup" and "blue cup" are close together, while "cup" and "plate" are moderately distant. These characteristics make them naturally suited for VLAs conditioned on natural language instructions.

SigLIP replaces CLIP [27]'s softmax contrastive loss with a sigmoid loss, reducing dependence on batch size and improving training efficiency. As of 2024–2025, the trend of SigLIP replacing CLIP [27] is pronounced.

• Advantages: Rich semantic alignment, optimal for language conditioning, benefits from large-scale pretraining
• Limitations: Lacks geometric precision. Tends to miss fine-grained spatial information such as "which direction does the cup handle face?"

4.1.2 Self-supervised Encoders

DINOv2 [30] is a ViT [28] encoder trained with masked image modeling and self-distillation. Because it learned from the structure of images themselves without language supervision, its representations are geometric. Object boundaries, surface normals, and spatial layouts are encoded with precision.

• Advantages: High geometric precision. Outperforms semantic encoders in contact-rich manipulation (pick-and-place, insertion, etc.). Captures subtle differences in texture and material.
• Limitations: No alignment with language, so an additional bridge is required for language-conditioned tasks such as "pick up the red cup."

4.1.3 Hybrid SigLIP + DINOv2 — The Current Dominant Standard

Since late 2024, hybrid encoding using both SigLIP and DINOv2 [30] simultaneously has become the de facto standard. OpenVLA [15]'s Prismatic VLM, OpenVLA [15]-OFT, GraspVLA [81], UniVLA [80], and others adopt this combination.

Why does this combination dominate? The answer lies in complementarity. SigLIP encodes "what," while DINOv2 [30] encodes "where and how." When processing the instruction "pick up the red cup," SigLIP contributes to identifying the semantic entity "red cup" in the scene, while DINOv2 [30] contributes to determining the orientation of the cup's handle and its precise location. The outputs of the two encoders are typically concatenated at the token level or integrated through a projection layer.

Empirically, the superiority of this combination has been confirmed repeatedly. In the Prismatic VLM paper, a comparison of SigLIP alone, DINOv2 [30] alone, and the SigLIP+DINOv2 [30] combination showed that the hybrid combination consistently outperformed on all benchmarks.

4.1.4 Using the Full VLM as an Encoder

Some models use an entire pretrained VLM as the encoder rather than a separate visual encoder. RT-H uses PaLI-X, π0 [16] uses PaliGemma, and VTLA uses Qwen-VL. The advantage of this approach is that the VLM already performs visual-language integration intrinsically, eliminating the need for a separate fusion module. The drawback is high computational cost, which is mitigated using efficient fine-tuning methods such as LoRA and QLoRA.

4.1.5 The Persistence of CNNs

Although ViT [28]-based encoders are mainstream, CNNs (ResNet, EfficientNet) have not disappeared. RT-1 [12] used EfficientNet-B3, and some lightweight models (LiteVLA [63], etc.) still choose CNNs for computational efficiency. In environments with extremely strict real-time constraints (1 kHz control loops in industrial robots) or on edge devices, the deterministic inference speed and small memory footprint of CNNs remain valid advantages.

4.1.6 Multimodal Perception

State-of-the-art VLAs integrate various sensory modalities beyond RGB cameras:

• Depth: SpatialVLA [39] encodes depth information as a separate channel to enhance 3D spatial understanding. Using the output of monocular depth estimation networks (MiDaS, Depth Anything) as additional input is also widely adopted.
• Tactile: ForceVLA [78] processes 6-axis force/torque sensor data, and TactileVLA processes GelSight tactile images alongside visual tokens. Tactile sensing is essential for delicate manipulation that requires "grasping firmly enough not to slip without gripping too hard."
• Force: OmniVTLA integrates vision, touch, and language in a single framework, encoding force profiles as temporal sequences.
• Audio: AudioCLIP [27]-based auditory encoders are explored experimentally. They have potential applications in auditory conditional behaviors such as "stop when you hear a click."

4.1.7 Proprioception Processing

Proprioceptive information such as the robot's joint angles, velocities, and end-effector position is typically converted into a fixed-dimensional vector through an MLP (Multi-Layer Perceptron). There are two main approaches for integrating this vector with visual-language representations:

• Concatenation: The proprioceptive vector is simply appended to the visual/language tokens. This is the straightforward approach adopted by most models, and is easy to implement.
• FiLM Conditioning: Feature-wise Linear Modulation, in which proprioceptive information modulates the scale and bias of visual features. Since proprioception directly influences visual processing, information integration is tighter. Octo [25] and HPT [96] use this approach.

---

4.2 Brain Module — The VLM Becomes the Robot's Brain

If the perception module is the "eyes," the brain module is the "central nervous system" that understands, reasons about, and plans from the perceived information. The VLA brain has undergone four stages of evolution from 2022 to 2025.

4.2.1 Four Stages of Evolution

Stage 1: Pure Transformer (2022–2023)

Gato [13] (DeepMind, 2022) was the first "generalist agent" to handle text, images, Atari games, and robot control with a single transformer. VIMA [45] proposed a transformer for understanding multimodal prompts, and GR-1 [51] proposed a transformer combining video generation and action prediction. The "brain" of this era was a transformer trained from scratch without general-purpose pretraining. Because it learned solely from robot data, it had fundamental limitations in language comprehension and visual commonsense reasoning.

Stage 2: Diffusion Transformer / DiT (2023–2024)

RDT-1B [24] (Robotics Diffusion Transformer) used a 1.2B-parameter DiT as the central architecture for action generation. TriVLA placed DiT as the core action generator in a triple system. DiT combined the scalability of transformers with the multimodal expressiveness of diffusion, enabling complex action distributions to be learned with large-scale models.

Stage 3: VLM + Generative Head (2024)

π0 [16] (Physical Intelligence, 2024) was the decisive turning point. It used PaliGemma (SigLIP + Gemma 2B) as the "brain" and attached a separate Flow Matching head as the "hand." (Note: PrismaticVLM, which combines SigLIP + DINOv2 + Gemma 2B, is the backbone of OpenVLA [15], not pi_0.) This was the first commercially successful instance of directly transferring a VLM's extensive visual-language knowledge to serve as a robot's brain. The key insight of this stage was: "There is no need to build a robot's brain from scratch — simply take a VLM already trained on internet-scale knowledge and connect a hand to it."

Stage 4: Fully VLM-based Brain (2024–2025)

The paradigm of "using a VLM directly as a robot policy," which began with RT-2 [11], matured through OpenVLA [15], π0.5 [31], CoT-VLA [55], and SafeVLA [75]. Models in this lineage directly leverage the VLM's language generation capability for action generation. CoT-VLA [55] goes one step further, explicitly outputting a reasoning process in natural language before generating actions. SafeVLA [75] internalizes safety constraints within the VLM's reasoning process.

4.2.2 Reasoning Paradigms

The VLA "brain" is evolving beyond simple reflex to reasoning.

Chain-of-Thought (CoT) Reasoning: A reasoning process is output in natural language before generating actions. ECoT [92] (Embodied Chain-of-Thought) generates reasoning chains such as "1. The red cup is on the left side of the table. 2. The gripper is currently on the right side of the table. 3. I must first move to the left." before outputting an action. CoT-VLA [55] trained this paradigm at scale and empirically demonstrated that reasoning improves action performance.

At ICLR 2026, Embodied Chain-of-Thought (ECoT [92]) emerged as a major trend. ACTIONS AS LANGUAGE [98], InstructVLA, and EMBODIED-R1, among others, integrate spatially-grounded reasoning with action prediction, moving beyond purely textual reasoning to introduce reasoning processes directly grounded in the visual scene into VLAs.

ReAct Paradigm: Reasoning and Acting are performed in alternation. The iterative loop of "observation → reasoning → action → observation → ..." allows environmental feedback to be incorporated into reasoning.

Visual Subgoal Prediction: Instead of language, future images are predicted to visually "imagine" what state should come next. SuSIE [59] and UniPi [93] explored this approach.

4.2.3 World Model Integration

World model integration endows the VLA's brain with "imagination." Two directions exist:

Policy Enhancement: Future predictions generated by the world model are used as auxiliary data or additional inputs for policy learning. UniVLA [80] predicts future states in latent space, and these predictions guide action generation. In WorldVLA [76], the video prediction module is jointly trained with the policy network.

Explicit Planning: The world model is used to simulate multiple possible futures, and the action leading to the most favorable future is selected. LUMOS [94] performs tree search with a latent space world model, and MinD [95] executes mental simulation within a latent world model.

The difference between these two directions lies in the role of the world model. In policy enhancement, the world model is an "advisor"; in explicit planning, the world model is a "simulator." The current trend is convergence of the two — toward a flexible architecture in which world model predictions directly intervene in policy action generation, while explicit simulation-based planning remains available for complex situations.

This world model integration can be situated within a broader decision-making framework. The Large Model Embodied AI survey [44] divides large-model-based embodied AI into hierarchical and end-to-end decision-making paradigms. In the hierarchical approach, high-level planning (VLM/LLM) and low-level control (specialist policies) are separated, yielding higher interpretability and safety but incurring inter-module information loss. In the end-to-end approach, a single model directly maps from perception to action, maximizing expressiveness but complicating debugging and safety assurance. The world model serves as a third axis that bridges these two paradigms — by endowing end-to-end models with plan-through-imagination capabilities, it absorbs the strengths of hierarchical architectures (safety, long-horizon reasoning) while preserving the benefits of a unified representation space.

---

4.3 Action Module — From Intention to Movement

After the brain decides "I should pick up the red cup," the action module's role is to translate this intention into actual motor commands. The fundamental challenge here is how to effectively represent and generate a continuous, high-dimensional action space.

4.3.1 Discrete Tokenization — The RT-2 Approach

This is the approach pioneered by RT-2 [11]. Continuous action values (e.g., joint angle 0.732 rad) are quantized into integer bins between 0 and 255, added to the VLM's vocabulary, and treated identically to language tokens. For a 7-DoF robot arm, the action at each timestep is represented as 7 tokens (plus 1 for gripper open/close).

• Advantages: Reuses the existing infrastructure of language models (tokenizer, generation algorithms, KV cache) as-is. Implementation is intuitive and simple. Since language generation and action generation share the same decoding process, multitask learning is natural.
• Disadvantages: 256-bin quantization limits the precision of the action space (1/256 ≈ 0.4% resolution). Multimodal action distributions cannot be represented — the model can only output a single mode, resulting in an averaging problem where a situation with two valid choices (rotate left or rotate right) yields "no rotation" as output. The autoregressive nature introduces latency proportional to the number of tokens.

4.3.2 Diffusion Policy — Stochastic Action Generation

DDPM (Denoising Diffusion Probabilistic Model): The approach pioneered by Diffusion Policy [17] (Chi et al., 2023), generating action sequences through iterative denoising starting from pure Gaussian noise. While 50–100 denoising steps are required, it faithfully represents multimodal distributions.

DDIM (Denoising Diffusion Implicit Model): The approach adopted by CogACT [23], which reduces denoising steps to 10–20 through a deterministic sampling process. It provides a trade-off between quality and speed.

Flow Matching: The approach adopted by π0 [16], modeling the path from noise to data as an ODE (ordinary differential equation). Training is more stable than DDPM/DDIM, and high-quality samples are generated in 5–10 steps. Rectified Flow learns approximately straight paths, further reducing the number of steps.

The common advantage of diffusion-based methods is multimodal distribution representation. They can simultaneously represent both modes of "the cup can be rotated left or right," sampling one at execution time. They naturally generate smooth trajectories and operate directly in continuous space with no quantization error.

4.3.3 FAST Tokenization — Innovation in the Frequency Domain

FAST [20] (Fast Action Tokenization) is an attempt to simultaneously achieve the simplicity of the autoregressive approach and the precision of continuous action representation. The core idea is to transform continuous action chunks into the frequency domain before tokenizing.

The specific process is as follows:

The results of this approach are impressive. A 5× pretraining acceleration compared to the RT-2 [11] approach is achieved with negligible information loss. This is because compression in the frequency domain is far more efficient than quantization in the time domain. The significantly reduced token count also alleviates the speed problem of autoregressive generation.

4.3.4 Normalizing Flows — The Dream of a Single Step

NinA (Neural Inference for Actions) applies Normalizing Flows to action generation. Unlike diffusion, distributions are modeled as a composition of invertible transformations, enabling sampling in a single forward pass. Since no multiple denoising steps are required, inference speed is very fast.

4.3.5 Frequency-domain Flow Matching — FreqPolicy

FreqPolicy performs Flow Matching in the frequency domain to achieve single-step inference. Action sequences are decomposed into frequency components, and a flow is learned in frequency space. Because complex multimodal distributions in the time domain have simpler structure in the frequency domain, sufficient-quality samples can be generated with fewer steps.

4.3.6 Action Chunking — Separating Temporal Scales

Action Chunking is a strategy that separates the temporal scales of action generation into a semantic level and a motor level. At the high level (low frequency), the "semantic direction of the next chunk" is decided autoregressively; at the low level (high frequency), "the fine-grained trajectory within the chunk" is generated in parallel.

ACT (Action Chunking with Transformers) popularized this concept. By generating action sequences of 16–32 timesteps in chunk units with a single prediction, it alleviates the myopic action problem of single-step prediction. Subsequent research has extended this to smoothness of inter-chunk transitions, variable-length chunks, and hierarchical chunking.

---

4.4 Dual-System Architecture — When Cognitive Science Meets Robotics

4.4.1 Kahneman's Dual-Process Theory

The dual-process theory presented by Daniel Kahneman in Thinking, Fast and Slow (2011) argues that human thought is composed of two systems:

• System 1: Fast, automatic, low-effort intuitive thinking. "A ball comes flying and you reach out your hand to catch it."
• System 2: Slow, conscious, high-effort analytical thinking. "Calculating the next move in chess."

The application of this theory to robotics is intuitively compelling. Robots also require two kinds of "thinking":

• Fast reflex (10–120 Hz): Motor control for avoiding obstacles, quickly re-grasping slipping objects, and maintaining smooth trajectories
• Slow deliberation (1–10 Hz): High-level reasoning for determining "which object to pick up?", "in what order to perform the task?", and "is this situation safe?"

The key insight is that these two systems operate on different time scales. System 2 does not need to run every frame, and System 1 does not need to perform complex reasoning.

4.4.2 Implementing the Robotic Dual System

System 1 — Action Expert: Handled by fast generative models such as diffusion policies, Flow Matching, and lightweight MLPs. Runs at 10–120 Hz to generate smooth and responsive motor commands. This module generates actions directly from a given context embedding, without the heavy reasoning of the VLM.

System 2 — VLM-based Reasoning/Planning: Handled by large-scale VLMs. Runs at 1–10 Hz to understand scenes, decompose tasks, and verify safety conditions. The output of this module is passed to the action expert as context — a directive of "what to do."

Asynchronous Execution: The core of the two systems is that they run at independent frequencies. While System 2 is reasoning about the next plan, System 1 continues to generate actions based on the previous plan. This asynchrony makes the slow inference speed of VLMs compatible with real-time control.

4.4.3 Implementation Examples

GR00T N1 [21] (NVIDIA): The Eagle-2 VLM (System 2) processes camera images and language instructions to produce context embeddings. These embeddings are passed to the DiT-based action expert (System 1), which generates action chunks via Flow Matching. This is a classic cascade structure following sequential information flow from System 2 → System 1.

π0 [16] (Physical Intelligence): The tokens from the PaliGemma VLM (System 2) and the tokens from the Flow Matching action expert (System 1) are processed simultaneously in shared attention blocks. In this parallel structure, the two systems share the same transformer layers and have access to each other's information.

MinD [95]: A latent world model plays the role of System 2, "mentally simulating" future states. The action policy (System 1) generates actions based on the results of this simulation. What is distinctive here is that System 2 is not linguistic reasoning but predictive simulation in latent space.

TriVLA: Proposes a triple system. The VLM (System 3, slowest) handles strategic planning, a DiT (System 2, intermediate) handles tactical action generation, and a lightweight executor (System 1, fastest) handles real-time correction. This extends the dual system to a triple system for finer-grained separation of time scales.

Hume: Targets holistic embodiment understanding, with the VLM and action generator processing a shared representation of whole-body human motion understanding.

4.4.4 Key Design Choice: Shared Attention vs. Cascade

The most important design choice in a dual-system architecture is the direction of information flow between the two systems.

Cascade (GR00T N1 [21] approach):

• Information flows unidirectionally from System 2 → System 1
• System 1 cannot provide feedback to System 2
• Advantage: Clear module separation allows each system to be trained/replaced independently. The VLM of System 2 can be upgraded while keeping the action expert of System 1 intact.
• Disadvantage: Since information from the action generation process is not reflected in reasoning, System 2 cannot respond to subtle changes that occur during action execution.

Shared Attention (π0 [16] approach):

• Information flows bidirectionally between System 1 ↔ System 2
• Language/visual tokens and action tokens interact within the same attention mechanism
• Advantage: The two systems can reference each other's state, enabling close cooperation in which action generation influences reasoning and reasoning guides action.
• Disadvantage: Unclear module separation means that replacing one system may require retraining the other. The computational cost of shared attention is high.

This choice reflects a fundamental trade-off between modularity and integration. Cascade provides the flexibility to swap components like Lego blocks, while shared attention creates a tightly integrated system like an organism.

Looking at the current (as of 2025) trend, the research community sees both approaches coexisting; however, models that have achieved commercial success (π0 [16], π0.5 [31]) lean toward shared attention, while models prioritizing platform/module replaceability (GR00T N1 [21]) lean toward cascade. Which will ultimately prevail remains an open question. What is clear, however, is that the cognitive science insight of separating "fast reflex" from "slow deliberation" has become a core principle of VLA architecture design.

---

4.5 Three Frontier VLAs, Honestly Compared — NVIDIA GR00T, Google Gemini Robotics, Physical Intelligence π

"The diagnosis that the three are variants of the same paradigm is correct. But if we stop there, we cannot see what to research next. We have to go one layer deeper."

When PhD students first start reading VLA papers in earnest, almost everyone follows the same trajectory. At first, GR00T [21], Gemini Robotics, and π [16] look like three different animals, because the demo videos from each company look so different. NVIDIA's humanoid moving objects, Google's videos of complex linguistic reasoning, PI's 13-hour espresso-making — they give wildly different visual impressions. Then, after reading Sections 3–6 of this survey to the end, the opposite recognition arrives: "aren't these all the same thing?"

Both intuitions are partly right and partly wrong. The purpose of this section is to find the precise point between them. The most dangerous thing for a first- or second-year PhD student entering this field is to settle at one of the two extremes. Viewing the three as "all different" makes you chase surfaces; viewing them as "all the same" makes you lose track of which current your research actually sits on. We have to honestly separate what is the same and what is genuinely different.

4.5.1 What to concede first: the unified-field diagnosis is 70% correct

If you have tracked the VLA field closely for the past 1–2 years, it is hard to disagree with the following proposition.

All three models are variants of a single paradigm: "an internet-pretrained VLM as the brain, a generative action decoder mounted on top, and the two modules executed asynchronously at different time scales."

This is exactly what this survey's Insight 1 (Evidence of Convergence) identifies, and forcing the three onto the five-axis coordinate system of Section 3.6 makes the point even sharper.

Gemini Robotics is not included in this survey's official five-axis classification table in §3.6. The row above is this section's inferred position based on public technical reports, while the rows for GR00T N1 and π0 match the §3.6 table exactly.

The differences in the table exist, but they are positional differences within the same category, not category-level differences. All three are dual-system, all three use generative action heads, and all three stand on pretrained VLM backbones. The four axioms identified by this survey (inheriting internet knowledge, time-scale separation, generative decoder, data pyramid) are genuinely the common ground of all three.

Any comparison that refuses to concede this ends up being a reshuffling of marketing copy. The difference among the three is not "different species" but "different breeds of the same species" — the conclusion of the unified-field argument must be accepted. Does the section end here? No. For a PhD researcher, the genuinely interesting question starts only after this.

4.5.2 What must not be flattened: intra-species differences are real too

If you tell a biologist "golden retrievers and Siberian huskies are the same species," they will agree. And yet their entire career is spent on the differences between those two breeds. Being in the same category does not mean the differences are meaningless. The paradigm being shared does not imply that divergence on top of that paradigm disappears. It was only after cars converged on "internal combustion + transmission + chassis" for 100 years that the real differences between Toyota, BMW, and Porsche became visible.

To see where the real differences in VLA lie, we must introduce three transverse axes that this survey, organized along technical axes, did not make explicit. These are precisely the points the unified-field argument intentionally flattened, and they are decisive when a PhD researcher decides which current to position their own work on.

Axis 1 — Data source determines the ceiling of generalization

What §10.12 of this survey explicitly pointed out is that even though frontier models and open-weight models are converging on simulation benchmarks (LIBERO, CALVIN), the gap in real-world zero-shot generalization is not closing. The root cause Reuss (2026)'s ICLR analysis points to is a gap in data curation and infrastructure scale. And this gap is not coincidence — it flows from the structural difference in where each of the three groups gets its data from.

NVIDIA centers on simulation synthesis via Cosmos and Isaac Sim/Lab. Simulation is infinitely scalable but pays a permanent tax called the sim-to-real gap. Google treats internet-scale multimodal pretraining as its asset. This is overwhelming for zero-shot semantic understanding, but it is a different dimension from the precision of motor control. PI relies on real-world data collected by its own robot fleet. This is the cleanest signal — no sim-to-real gap — but it is the most scalability-constrained.

Which data strategy a PhD researcher imitates in their own lab is decisive. If you have simulation infrastructure, following the NVIDIA line is natural; if not, fine-tuning openpi (which PI has released) with a small amount of real-world data is the realistic path. This choice is not merely a tool choice — it is a choice of what you accept as the upper bound of generalization.

Axis 2 — Which training stage you pour resources into

Recall the three-stage maturity model from §6.6 (internet pretraining → BC fine-tuning → RL post-training). The interesting fact is that each of the three groups has picked a different stage as its point of differentiation. And this is not coincidence — it is determined by each group's asset structure.

Google's assets are overwhelmingly concentrated on Stage 1 (internet pretraining): massive data centers, internet-scale multimodal corpora, TPU infrastructure. So Gemini Robotics' differentiation becomes "transplanting the enormous Gemini backbone as-is." NVIDIA follows the orthodox path on Stages 1–2 at the model level (inheriting the Eagle-2 VLM + BC fine-tuning), but differentiates the data pipeline that feeds Stages 1–2 via Cosmos/Isaac Sim. In other words, its differentiation is the data pipeline, not the model recipe. PI takes Google's open PaliGemma/Gemma 3 for Stages 1–2, and defines Stage 3 (RL post-training) and deployment-time learning beyond it as its own territory. π^*_0.6 [157]'s advantage conditioning (§10.13.1), which first demonstrated a practical path for applying RL to Flow Matching VLAs, and π_0.6-MEM [158] (§10.13.2), which solved 15-minute long-horizon tasks with multi-scale memory, both occur within this Stage-3 territory.

For a PhD researcher, this translates into a very practical question. At which training stage can you contribute over the next 3–5 years? Competing with massive backbones on Stage 1 is effectively impossible for a single academic lab. Stage 2 BC fine-tuning is already well-paved. The most open frontier is Stage 3 RL post-training and its variants — the territory PI is pushing fastest, but also the territory with the lowest barrier to entry. The trajectory of self-improving residual RL reaching 99% on LIBERO at ICLR 2026 supports this diagnosis.

Axis 3 — Model-weight release policy creates an asymmetric ecosystem

This is a political-economic dimension the survey does not directly address, but it is the difference that most directly affects the daily life of a PhD researcher. NVIDIA releases GR00T N1 on Hugging Face and elsewhere; openpi releases π0/π0-FAST but everything after π0.5 is closed; Gemini Robotics is an API/partnership model from the start.

This difference is not merely corporate policy — it determines who can do what kind of follow-up research. Attempting ablation studies or mechanism interpretation in academia without weight access is nearly impossible. So the de facto baselines of academic VLA research are open models like OpenVLA [15], π0, and GR00T N1; Gemini Robotics is a citation target, not a comparison target. This means the "frontier vs. open-weight gap" in §10.12 is not merely a performance gap but a gap in the very possibility of research.

A first- or second-year PhD student should be explicitly conscious of this when choosing a new topic. "Research that compares against Gemini Robotics" will almost always be confined to limited forms (citing published numbers, comparing API calls). In contrast, "research that verifies a new RL post-training method on top of openpi" is immediately executable and reproducible by others.

4.5.3 So, the real coordinates of the three

Combining the three axes, the differences among the three can be honestly summarized in one line.

Note the relationship between this table and the table in 4.5.1. If 4.5.1 shows convergence along the technical axis, this table shows divergence along the strategic/ecosystem axis that sits on top of it. Both tables are true; looking at only one misreads the field as a whole.

4.5.4 From the PhD researcher's view — what to track and what not to follow

Rather than ending this section on a neutral, compromising note, it is more honest to close with concrete guidance for the decisions a first- or second-year PhD student will actually face.

Areas where you must track all three. Evolution of dual-system architectures, the mathematical mechanics of action decoders (diffusion vs. flow matching vs. discrete diffusion), VLM backbone choice and fine-tuning strategies. In this area, all three are genuinely playing the same game, and progress in one group foreshadows the near future of the others. Sections 4–5 of this survey correspond to this area.

Areas to track by group, divergently. Data collection/augmentation strategies (NVIDIA's Cosmos current, Google's internet-multimodal current, PI's real-world-fleet current), post-training-stage innovations (especially PI's RL post-training and memory integration), domain-specialized applications (NVIDIA's humanoid line, Google's EMMA autonomous-driving lineage). In this area, group-level asset structures differ, so a result from one group is not easy to transplant onto another.

Areas where tracking only one group is enough. PI's π^*_0.6 and π_0.6-MEM lineage. The reason this survey devotes all of §10.13 to two subsections on these papers is that this current is currently breaking through the field's most urgent open problems (BC's performance ceiling, the absence of memory in long-horizon tasks) head-on. If a PhD researcher is interested in RL post-training or memory mechanisms, starting from openpi and the π-series follow-ups as baselines is the most efficient path.

Traps to consciously avoid. The question "which company is ahead" — driven by the impressions of corporate demo videos — is a question that academia cannot answer. Another common trap is tracking only simulation-benchmark numbers and missing the real-world generalization gap (§10.12). And the biggest trap — settling on the unified-field argument too early and dismissing group-level differences with "they're all the same paradigm anyway." That the paradigm is shared is not a license to ignore divergence — it is a tool that lets you predict where divergence will occur more precisely.

4.5.5 One-line summary

The three groups are variants of a single paradigm at the level of mathematical architecture, and that diagnosis must be taken seriously. But along the transverse axes of data source, training-stage resource allocation, and model release policy, they genuinely sit at different coordinates, and that difference is what decides which current a PhD researcher places their own research on. In five years, the models themselves are more likely to resemble one another even more closely; but the ecosystems they produce and their relationships to academia are more likely to diverge further. Seeing both currents at the same time is the balance a first- or second-year PhD student needs when settling into this field.

---

Chapter 4 Summary: Current Coordinates of VLA Anatomy

The perception module is converging on SigLIP + DINOv2 [30] hybrid as the dominant standard. The brain module has undergone definitive evolution toward directly repurposing pretrained VLMs as the robot's brain. The action module has no definitive winner yet, with discrete tokenization, diffusion, Flow Matching, and FAST tokenization [20] all in competition. And as the approach for connecting these three modules, the dual-system architecture is rising in prominence, with insights from cognitive science being translated into engineering design principles.

The next chapter examines how to train the architectures constructed in this way — delving deeply into the three paradigms of behavioral cloning, reinforcement learning, and world model learning.

---

Motivation Chain: The Causal Logic of Architectural Evolution

Limitations of modular pipelines (sense-plan-act separation leads to information loss and engineering overhead)
--> Monolithic architecture emerges (RT-2 [11]: a single VLM handles everything)
--> Limitations of monolithic models (VLM inference is too slow for real-time control)
--> Dual-system architecture emerges (GR00T N1 [21]: fast System 1 + slow System 2)
--> Limitations of dual systems (insufficient planning capability for long-horizon compound tasks)
--> Hierarchical architecture (pi_0.5: VLM planner + VLA executor)

Limitations of autoregressive decoding (single-mode generation only, discretization information loss, slow sequential generation)
--> Diffusion Policy [17] emerges (multimodal action representation, batch action chunk generation)
--> Limitations of DDPM (50--100 denoising steps make generation slow)
--> Flow Matching (pi_0 [16]) emerges (linear interpolation converges in 5--10 steps)
--> DiT-based decoders (CogACT [23], RDT-1B [24]) emerge (applying transformer scaling laws to diffusion)

---

Discriminative Features of Similar Architectures

---

Intuitive One-Liners: Architecture Edition

• Monolithic VLA: "A person who understands a foreign language and responds directly without an interpreter -- fast, but deep deliberation is difficult."
• Dual-system: "A CEO (strategic decisions, slow) and a shop-floor worker (immediate execution, fast) dividing labor."
• Hierarchical architecture: "Separating the recipe (high-level plan) from knife skills (low-level technique) -- changing the recipe does not require relearning how to cut."
• Autoregressive decoding: "Typing one character at a time -- actions are generated one token after another in sequence."
• Diffusion decoding: "Sculpting from marble -- the action emerges as unnecessary material (noise) is chiseled away from the whole."
• Flow Matching: "A shortcut through diffusion's 'sculpting' -- reaching the same result in fewer steps via a straight-line path."
• Action Chunking: "Generating an entire sentence at once rather than one character at a time -- securing temporal coherence."

---

Self-Check Questions: Sections 3--4

Q1: How is π0 [16]'s architecture classified under Liu & Shao's taxonomy and under Zhong et al.'s taxonomy, respectively?

Answer: Under Liu & Shao's taxonomy, π0 [16] is a Monolithic Parallel Dual-system -- the PaliGemma VLM (System 2) and the Flow Matching Action Expert (System 1) process tokens simultaneously within shared attention blocks. Under Zhong et al.'s taxonomy, π0 [16] falls under Diffusion-based (Flow Matching) action generation, since Flow Matching is a variant of the diffusion family.

Q2: Why does the autoregressive approach (e.g., RT-2 [11]) struggle to represent multimodal action distributions?

Answer: The autoregressive approach quantizes each action dimension into discrete bins and generates tokens sequentially. This process incurs two problems: (1) quantization error degrades the precision of continuous actions, and (2) because each token is conditioned sequentially on the preceding tokens, the model can only represent a single mode of the distribution. When two equally valid modes exist (e.g., "grasp from the right" and "grasp from above"), the model tends to converge on the average direction (mode averaging), outputting an action that corresponds to neither valid option.

Q3: Among Xu et al.'s Five Challenges, which has seen the greatest progress and which remains the most unresolved as of 2026?

Answer: The greatest progress has been in (1) Representation -- VLM-based unified representations, FAST tokenization [20], and Flow Matching have substantially improved the representation problem. The most unresolved challenge is (4) Safety -- SafeVLA [75] represents an initial attempt, but formal verification is absent, and the fundamental risk that VLA hallucinations can lead to physical accidents remains unaddressed.

---

Open Research Questions: Sections 3--4

5. Action Tokenization — The Core Design Decision in VLA

When designing a Vision-Language-Action (VLA) model, countless architectural choices exist, but the most fundamental and far-reaching decision is "how to represent actions as tokens." Chen et al. [7] (2025) termed this the Action Tokenization perspective and presented it as the axis that most clearly differentiates VLA models from one another. This section systematically organizes 8 action token types based on their classification, integrating insights from other surveys.

Why is tokenization central? VLA models are fundamentally built on large language models (LLMs). Since LLMs are systems that process discrete token sequences as input and output, the method by which a robot's continuous action space is converted into a discrete token space fundamentally determines the model's expressive power, control precision, and inference speed. The choice of tokenization method is not merely an implementation detail — it is the highest-level design decision that defines the capabilities and limitations of the VLA system.

A single model can combine multiple token types; for example, CoT-VLA [55] jointly utilizes reasoning tokens (Type 8) and goal tokens (Type 5).

---

5.1 Eight Action Token Types

Type 1: Language Tokens

The most intuitive approach is to represent robot actions as natural language text. An LLM generates a natural language command such as "pick up the red cup," and a low-level policy or predefined skill primitive then converts this into actual motor commands.

Representative models:

• SayCan [14] (Ahn et al., 2022): Scores action candidates generated by the LLM with affordance scores to select executable actions. Introduced the key idea of combining the LLM's world knowledge with the robot's physical capabilities via a product operation.
• Inner Monologue [22] (Huang et al., 2023): Converts feedback from the environment (success/failure, object recognition results) into language and incorporates it into the LLM's next action plan.
• SayTap [38] (Tang et al., 2023): Represents a walking robot's foot contact patterns as text sequences, enabling the LLM to directly plan locomotion rhythms.

Advantages: The LLM's powerful language generation capability and commonsense reasoning can be directly leveraged. Because pre-trained linguistic knowledge transfers intact, zero-shot generalization to new tasks is excellent.

Limitations: Control precision for continuous motor control is inherently insufficient. It is difficult to adequately express fine-grained manipulation such as "move the cup 3 cm to the left" in natural language. The control frequency is 1–3 Hz — the lowest among all token types — making this approach unsuitable for tasks requiring rapid response (e.g., catching a moving object).

---

Type 2: Code Tokens

This approach represents actions as executable program code. The LLM generates Python functions or API call sequences, which are then executed directly on the robot runtime.

Representative models:

• Code as Policies [36] (Liang et al., 2023): The LLM directly generates Python code that calls robot APIs. Complex action sequences can be composed by leveraging programming constructs such as spatial reasoning, loops, and conditionals.
• ProgPrompt [102] (Singh et al., 2023): Structures tasks in a programmatic format (function calls, assert statements) to enhance the LLM's planning capability.
• Voyager [103] (Wang et al., 2023): Generates executable code in the Minecraft environment and stores successful code in a skill library, progressively expanding capability.
• ChatGPT for Robotics [104] (Vemprala et al., 2024): Converts user intent into robot control code through a conversational interface.

Advantages: Structured, reusable, and easy to debug. Complex action logic can be expressed concisely via loops and conditionals, and generated code can be accumulated in a library to expand the skill repertoire over time.

Limitations: Predefined APIs (e.g., pick(obj), place(x, y, z)) are required, and fine-grained dexterous manipulation not supported by the API cannot be expressed. The continuous nature of physical interaction is difficult to capture fully through discrete API calls.

---

Type 3: Affordance Tokens

This approach spatially represents the manipulable regions and modes of objects. "Where" to grasp and "in which direction" to push are expressed as heatmaps or vector fields in 3D space.

Representative models:

• VoxPoser [57] (Huang et al., 2023): Uses an LLM and VLM to generate affordance maps (value maps) and constraint maps in a 3D voxel space. A motion planner synthesizes trajectories based on these maps.
• A3VLM [58] (Huang et al., 2024): Extends a VLM to directly predict affordances from 3D point clouds.
• RT-Affordance [105] (Brohan et al., 2023): Uses visual affordances as a conditioning signal to aid generalization of the manipulation policy.
• A0 [106] (Ren et al., 2025): A unified affordance-based manipulation framework that explicitly models object–action relationships.

Core value: Affordance tokens provide an intermediate representation of "where" and "how" to manipulate. They serve as a semantic bridge between high-level language planning and low-level motor control, and are particularly strong at generalizing to novel objects.

---

Type 4: Trajectory Tokens

This approach represents the spatiotemporal path of an end-effector as a token sequence. Future trajectories are sketched onto 2D images, or waypoint sequences in 3D space are generated.

Representative models:

• RT-Trajectory [62] (Ahn et al., 2024): Overlays 2D trajectory sketches onto images for use as visual prompts. Humans can draw trajectories, or the model can predict them.
• LATTE (Liu et al., 2024): A language-to-trajectory converter that transforms language commands into 3D trajectory sequences.
• TraceVLA [40] (Zheng et al., 2025): The VLA model generates visual trajectory traces as an intermediate representation, simultaneously improving the interpretability and accuracy of action prediction.

Connection to video pre-training: Trajectory tokens connect naturally with video prediction models. Since a video frame sequence is essentially a continuous visual trajectory, the spatiotemporal understanding learned by models pre-trained on large-scale video data can be directly transferred to trajectory prediction.

---

Type 5: Goal Tokens

This approach represents goal states by predicting future observations. "How will the world look after an action from the current state?" is generated as an image or point cloud.

Representative models:

• SuSIE [59] (Black et al., 2024): Takes the current observation and a language command as input, generates future subgoal images, and trains a low-level policy to follow them.
• UniPi [93] (Du et al., 2024): Frames robot planning as a video generation problem. A text-to-video diffusion model generates a future frame sequence, and an inverse dynamics model extracts the actions.
• 3D-VLA [37] (Zhen et al., 2024): Predicts future scenes in a 3D representation space, generating goals that reflect rich spatial understanding.
• CoT-VLA [55] (Kang et al., 2025): Uses visual subgoals as a Chain-of-Thought, explicitly reasoning about "how the world should look next" before deciding on an action.

Integration with world models: Goal tokens offer the most natural interface with world models. Setting a goal by simulating the future is fundamentally equivalent to imagining the consequences of an action through an internal world model. This presents a powerful pathway for integrating planning and execution.

---

Type 6: Latent Tokens

This approach represents actions in a learned latent space. Raw action data is compressed with an autoencoder or similar mechanism, converting it into semantically rich latent vectors.

Representative models:

• LAPA [61] (Ye et al., 2025): Uses a VQ-VAE (Vector Quantized Variational Autoencoder) to quantize action trajectories into a latent codebook. The VLA model is trained to predict these latent action tokens.
• UniVLA [80] (Li et al., 2025): Learns a unified latent action space to handle diverse robot embodiments and tasks within a single model.
• VQ-VLA [107] (Qu et al., 2025): Discretizes the action space via vector quantization while preserving the structure of the latent space to maintain semantic coherence.

Key to bridging the embodiment gap: The most innovative value of latent tokens lies in overcoming the embodiment gap. Although human hand movements and robot gripper movements are physically entirely different, the semantic action of "grasping an object" can be represented similarly in the latent space. This makes it possible to transfer action knowledge extracted from large-scale human video data (Ego4D, Something-Something, etc.) to robots — a key strategy for circumventing the chronic shortage of robot data.

Domain-agnostic properties: Because the latent action space does not depend on the joint configuration or action-space dimensionality of a specific robot, it becomes a natural medium for cross-embodiment learning.

---

Type 7: Raw Action Tokens

This approach directly discretizes low-level action values — such as joint angles, end-effector pose (position + orientation), and gripper state — and converts them into tokens. It is the most direct and simple tokenization method.

Representative models:

• RT-2 [11] (Brohan et al., 2023): Each dimension of a 7-dimensional action vector (6-DoF pose + gripper) is uniformly discretized into 256 bins and added to the LLM's vocabulary.
• OpenVLA [15] (Kim et al., 2024): Adopts the same 256-bin discretization as RT-2 [11], implemented in an open-source, reproducible manner.
• Gato [13] (Reed et al., 2022): Discretizes actions from diverse tasks into 1024 bins and processes them with a single general-purpose model.

The problem of quantization error: The fundamental limitation of raw action discretization is quantization error. Discretizing into 256 bins gives each bin a width of approximately 0.8%; in precision manipulation, this error can accumulate and cause failures. Moreover, tokenizing each action dimension independently loses inter-dimensional correlations.

The innovation of FAST: FAST [20] (Fast Action Tokenization) (Pertsch et al., 2025) proposed an elegant solution to this problem. It applies a Discrete Cosine Transform (DCT) to action sequences to convert them to the frequency domain, then tokenizes them using Byte Pair Encoding (BPE). This achieves:

• Improved information compression as the DCT captures temporal correlations
• Shorter sequence lengths as BPE groups frequent action patterns into single tokens
• Natural integration with the LLM's existing vocabulary expansion mechanism

This innovation dramatically improves the efficiency of raw action tokenization, reducing the number of tokens by up to ~13x compared to the conventional 256-bin method at equivalent precision.

At ICLR 2026, next-generation tokenization techniques surpassing FAST were proposed. FASTer combines Residual Vector Quantization (RVQ) with frequency- and time-domain losses to simultaneously achieve higher compression ratios and superior reconstruction quality. OMNISAT employs a B-Spline encoder to provide a compact representation specialized for smooth, long-horizon action outputs.

---

Type 8: Reasoning Tokens

This approach generates the reasoning process as explicit tokens before deciding on an action. The model first reasons about "why this action should be taken," and then predicts the action based on that reasoning.

Representative models:

• ECoT [92] (Zawalski et al., 2024): Embodied Chain-of-Thought. Generates text describing scene descriptions, task analysis, and sub-plans before predicting actions.
• CoT-VLA [55] (Kang et al., 2025): Combines visual reasoning tokens (future subgoal images) with linguistic reasoning tokens to construct a multimodal chain of thought.
• ThinkAct [67] (Xu et al. [8], 2025): Introduces an adaptive reasoning mechanism that autonomously determines "when to think."
• Embodied-R1 [108] (Liu et al., 2025): Applies DeepSeek-R1-style long-chain reasoning to embodied tasks, generating spontaneous reasoning paths for complex multi-step manipulation.

Performance improvement effects: The effect of reasoning tokens is substantial. According to experiments by SC-VLA [56], including reasoning tokens improved action prediction quality by approximately 35%. This demonstrates that "think before acting" provides a clear advantage over simple reactive behavior.

Trade-offs: Reasoning tokens require additional token generation, which increases inference latency. ThinkAct [67]'s approach of learning "when to think" is an attempt to manage this trade-off — acting immediately for simple tasks and activating reasoning only for complex situations.

---

5.2 The Relationship Between Tokenization and Control Frequency

The tokenization method directly determines the control bandwidth of a VLA model, which fundamentally defines the range of tasks the model can perform.

The key insight revealed by this table is clear: the tokenization method determines the control bandwidth, which defines the range of executable tasks. A control frequency of 1–3 Hz can handle simple tasks such as "put the red block in the blue bowl," but cannot achieve grasping an egg without crushing it. Only at frequencies above 50 Hz does delicate manipulation requiring force control become feasible.

The root cause of this frequency gap lies in the token generation mechanism:

• Autoregressive decoding generates tokens one by one sequentially, so latency increases proportionally with the number of action dimensions (7-DoF → 7 tokens generated sequentially).
• Diffusion/flow matching-based action chunking simultaneously generates dozens of steps of actions in a single denoising pass. Although the inference frequency is low, the actions within the chunk are executed at high frequency.
• FAST [20] uses DCT+BPE to compress action sequences into a small number of tokens, achieving a high effective frequency even with autoregressive decoding.

---

5.3 The Impact of Tokenization Choice on Performance

Discrete vs. Continuous: The Multimodal Distribution Problem

The most serious impact of the tokenization choice on performance appears in multimodal distribution scenarios. Consider, for example, a situation where an object on a table can be pushed in either direction — left or right. In this case, the correct action distribution has a bimodal form — probability mass concentrated at both left and right.

When trained with discrete raw tokens (256-bin) and standard cross-entropy loss, the model tends to predict the action corresponding to the average of the two modes (i.e., pushing in neither direction). This is the mode averaging (mode collapse) problem, which becomes more severe as the diversity of demonstration data increases.

Diffusion-based continuous action generation provides a natural solution to this problem. Diffusion models can inherently represent multimodal distributions, capturing both modes of the bimodal distribution. This is one of the core reasons π0 [16], GR00T N1 [21], and others have adopted diffusion/flow matching.

Co-determination of Action Space and Decoding

Tokenization methods and decoding strategies are not chosen independently — they are co-determined:

• Discrete action space ↔ autoregressive decoding: Representing actions as discrete tokens allows the LLM's existing language model head to be reused as-is. RT-2 [11] and OpenVLA [15] take this path. The advantage is that architectural modifications are minimized, maximally preserving the LLM's pre-trained knowledge.
• Continuous action space ↔ diffusion/flow-based generation head: Keeping actions as continuous vectors requires a dedicated generation head (diffusion head, flow matching head). π0 [16] and GR00T N1 [21] take this path. The advantages are multimodal distribution representation and high control frequency; the disadvantage is that additional design is needed for integration with the LLM backbone.

Recent work has also attempted to combine these two paths. For example, VQ-VLA converts continuous actions into discrete tokens via vector quantization while preserving the structure of the latent space, aiming to combine the advantages of both worlds.

Trade-offs of Action Chunking

Action chunking is a technique for simultaneously predicting actions across multiple timesteps in a single inference pass. Proposed in ACT (Zhao et al., 2023), it has since become a core element of VLA design.

Increasing the chunk size:

• Reduces inference frequency, improving computational efficiency (e.g., chunk size 16 → 1/16 the number of inferences)
• Improves trajectory consistency — predicting independently at each step can lead to an unstable trajectory, whereas chunk-level prediction guarantees temporal coherence
• However, reduces responsiveness to environmental changes — if an unexpected situation arises in the middle of chunk execution, immediate response is not possible

To manage this trade-off, π0 [16] and others adjust chunk size to match task characteristics, or apply temporal ensembling to smoothly interpolate (interpolation) actions between successive chunks.

---

5.4 Summary: An Integrated Understanding of the Tokenization Perspective

The 8 action token types can be understood as a spectrum of abstraction levels:

High abstraction  ←─────────────────────────────────→  Low abstraction
Language → Code → Reasoning → Goal → Trajectory → Affordance → Latent → Raw

Moving to the left yields greater human interpretability and generalization but lower control precision; moving to the right yields greater precision but weaker generalization and interpretability. The most successful modern VLA systems hierarchically combine multiple levels of this spectrum — for example, forming a plan with reasoning tokens (high abstraction) and then executing with raw action tokens (low abstraction).

From this perspective, the central question in VLA research is not "which token type is best?" but rather: "For which tasks and embodiments is which combination of token types optimal?"

The ordering above represents one perspective; the relative abstraction levels of affordances and trajectories may vary depending on the domain and task.

---

---

---

6. The Evolution of Learning Paradigms

The training of VLA models has evolved beyond simple supervised learning into a multi-layered process spanning from pre-training to post-training. This section systematically examines each stage of this evolution and explores a fascinating parallel with theories of human motor learning.

---

6.1 Pre-training — From the Internet to Robots

Two-phase joint training has become the de facto standard for VLA model pre-training.

Phase 1: Internet-Scale Image–Text Pre-training

In the first phase, a vision-language model (VLM) is trained on large-scale image–text data collected from the internet. Through datasets such as LAION-5B (5 billion image–text pairs), COCO, and Visual Genome, the model acquires a rich semantic prior about the visual world.

What the model learns at this stage is not robot control itself, but the world understanding that underlies it:

• Object recognition and classification ("this is a mug")
• Understanding of spatial relationships ("the mug is on the table")
• Reasoning about physical properties ("a glass can break")
• Commonsense action knowledge ("to drink from a mug, hold the handle")

Phase 2: Fine-tuning on Robot Trajectory Data

In the second phase, the model is fine-tuned on real robot trajectory data. Key datasets include:

• Open X-Embodiment [19] (OXE [19]): A cross-embodiment dataset containing 22 robot morphologies and more than 1M+ episodes. It is currently the de facto standard data source for VLA training.
• BridgeData V2: Manipulation data from a WidowX robot across diverse environments (~60,000 trajectories).
• RT-1 [12] data: A large-scale single-embodiment dataset collected by Google's Everyday Robots.

The core insight of this two-phase structure is the separation of semantic prior knowledge from sensorimotor skills. The ability to understand the world (Phase 1) and the ability to act in the world (Phase 2) can be efficiently learned from different data sources.

The Rise of Video Pre-training

Recently, the trend of utilizing video data in pre-training — beyond image–text data — has been accelerating:

• GR-2 [100] (Cheang et al., 2024): Uses a video generation model pre-trained on web-scale video as the foundation for a robot policy. Understanding of physical dynamics embedded in video transfers to robot control.
• Egocentric video (Ego4D, EPIC-Kitchens): First-person-view videos of humans performing direct manipulation are similar to the robot's perspective, providing particularly rich prior knowledge for manipulation tasks.

The decisive advantage of video pre-training over image–text pre-training is the understanding of temporal dynamics. Images provide understanding of static scenes, while video provides understanding of "how the world changes after this action."

The Role of Simulation Data

• UniSim [101] (Yang et al., 2023): An action-conditioned video diffusion model that generates unlimited training data through simulated interaction in virtual environments.
• Genesis (Xian et al., 2024): A GPU-accelerated physics simulator that generates large-scale, physically realistic interaction data.

The primary challenge with simulation data is the sim-to-real gap — visual and physical differences between simulation and the real world. Techniques such as domain randomization and domain adaptation are being actively researched to reduce this gap.

Scaling Laws

Scaling laws have also been observed in VLA training. According to research by Zhang et al. [6], doubling the amount of trajectory data improves task success rate by approximately 8--12% (original paper citation; specific figure not in surveys). This means that acquiring more data translates directly into performance improvement, underscoring the importance of large-scale data collection infrastructure (OXE [19], DROID, etc.).

However, data scaling alone has limits. The quality, diversity, and coverage of trajectory data are just as important as quantity, and developing more efficient learning algorithms must proceed in parallel with simply increasing the amount of data.

---

6.2 Limitations of Behavioral Cloning

Behavioral Cloning (BC) is the most basic policy learning method, which imitates expert demonstrations through supervised learning. Given observation–action pairs $(o_t, a_t)$, a policy $\pi(a|o)$ is trained via maximum likelihood estimation (MLE). The majority of VLA models are built on this BC framework.

However, BC has fundamental limitations:

Distribution Shift

The state distribution the model sees during training differs from the state distribution the model encounters during execution. During expert demonstration, the model follows the state distribution generated by the expert policy $\pi^*$, but during execution it follows the state distribution generated by the trained (imperfect) policy $\hat{\pi}$. This distributional mismatch leads to unpredictable behavior in states not seen in the training data.

Covariate Shift and Compounding Errors

Small prediction errors at each timestep accumulate over time. A slight positional error at one step leads to a larger error at the next step, which cascades and amplifies to cause catastrophic failures in long-horizon tasks. For example, in a complex assembly task lasting 30 seconds, a minute early error can result in complete failure at a later stage.

Inability to Improve Beyond Suboptimal Demonstrations

BC is inherently bounded by the upper bound of the demonstrations. If the demonstrations themselves are not optimal or contain noise, the model cannot surpass that level. There is no mechanism for discovering better actions.

Absence of Safety/Preference Signals

BC learns only "what to do" and cannot leverage signals about "what not to do" or "which actions are preferred." It cannot explicitly reflect safety constraints (e.g., speed limits near humans) or user preferences (e.g., preference for smooth motions).

Core Conclusion

BC is a necessary but not sufficient condition for VLA training. While BC is indispensable for efficiently leveraging large-scale demonstration data, the need for post-training to overcome its limitations is becoming increasingly clear.

---

6.3 Reinforcement Learning Post-training

Centered on Jin et al. [9] (2025), research on using reinforcement learning (RL) to push VLA model performance one step beyond BC has been growing explosively. This parallels precisely the paradigm in the large language model (LLM) field of aligning models with RLHF (Reinforcement Learning from Human Feedback) after SFT (Supervised Fine-Tuning).

Online Reinforcement Learning

The model learns by directly interacting with a real environment (or simulation) and receiving rewards:

• PPO-based:
• VLA-RL [68] (Tan et al., 2025): Applies PPO to a VLA model for performance improvement through online environment interaction
• RIPT-VLA [71] (Su et al., 2025): Reinforcement learning via Iterative Policy Training. According to the original paper (Su et al., 2025), on a specific task it starts from a 4% SFT success rate and reaches a 97% success rate after 15 PPO iterations. Note, however, that in Jin et al. [9]'s LIBERO benchmark comparison the average is 74.7%, indicating substantial variation across benchmarks.
• iRe-VLA (Xu et al. [8], 2025): Progressively improves the policy through iterative RL

• GRPO-based:
• ThinkAct [67] (Xu et al. [8], 2025): Applies Group Relative Policy Optimization to simultaneously reinforce reasoning and action
• TGRPO [66] (Li et al., 2025): An extended GRPO that includes reasoning consistency rewards

Offline Reinforcement Learning

Improves the policy using only previously collected data. Better actions can be extracted from suboptimal demonstrations without additional environment interaction:

• PA-RL: Selects optimal trajectories from offline data through CalQL (Calibrated Q-Learning)-based re-ranking
• ConRFT [69] (Li et al., 2025): Online reinforcement fine-tuning using a consistency policy

Preference Optimization

Directly uses human preferences as learning signals:

• HAPO [84] (Li et al., 2025): Applies DPO (Direct Preference Optimization) to VLA. Learns preferred action patterns through pairwise trajectory comparisons
• RAPL [83] (Tian et al., 2025): Learns a reward function through visual preference encoding, simply by having humans compare video clips
• GRAPE [109] (Wang et al., 2025): Multi-scale preference learning — simultaneously incorporates preferences at the trajectory level, segment level, and step level

The Spectrum of Reward Design

The central challenge of RL post-training is designing an appropriate reward function. Reward types proposed to date include:

Key Performance Figures

Impressive numbers demonstrating the effect of RL post-training:

• RIPT-VLA [71]: SFT 4% → 97% success rate after 15 PPO iterations on a specific task per the original paper (Su et al., 2025); Jin et al. [9] report a LIBERO average of 74.7%.
• SimpleVLA-RL [70]: 17.3% → 91.7% (with just 1 trajectory per task) (original paper citation; source outside the 14 surveys).
• These figures prove that RL post-training can deliver a fundamental performance leap, not merely incremental improvement.

Self-improving residual RL methods presented at ICLR 2026 have reached 99% success rates on LIBERO, demonstrating that the potential of RL post-training can push performance to benchmark-saturation levels. Stage-aware reinforcement learning is a novel approach that decomposes tasks into semantic constituents and optimizes each stage independently.

Resolving BC→RL Transition Instability

The greatest practical challenge when applying RL to a BC-initialized model is training instability. RL updates can destroy useful action patterns learned during BC (catastrophic unlearning). Strategies to address this:

• BC loss regularization: Adds BC loss as a regularization term to the RL objective, preserving the basic capabilities learned during BC.
• VL encoder freezing: Freezes the vision-language encoder weights and updates only the policy head with RL, preserving pre-trained visual-language understanding.
• Dual-Q/ensemble critics: Suppresses overestimation of the value function to secure training stability.

---

6.4 Parallels with Human Motor Learning

Jin et al. [9] (2025) noted that VLA learning paradigms have a structure remarkably similar to human motor learning theory. This analogy goes beyond mere metaphor and offers practical insights into the future direction of VLA research.

Newell's Constraints-Led Theory (1986)

Karl Newell argued that motor behavior emerges from the interaction of three types of constraints. This framework directly corresponds to VLA design:

Environmental Constraints:

• Human: gravity, friction, physical properties of objects, etc.
• VLA: affordance recognition, perceptual enhancement modules → encoding physical constraints of the environment into the model

Organismic Constraints:

• Human: body size, muscle strength, range of joint motion, etc.
• VLA: embodiment awareness → learning forward kinematics and inverse kinematics

Task Constraints:

• Human: task goals, rules, time limits, etc.
• VLA: hierarchical task decomposition, Chain-of-Thought reasoning → decomposing complex tasks into manageable subtasks

Neuroscientific Correspondences

Each component of VLA functionally corresponds to a specific system in the human brain:

Particularly noteworthy in this correspondence is the similarity between basal ganglia chunking and action chunking. The process by which humans automate complex motor sequences (e.g., piano playing) into a single "chunk" through repeated practice is strikingly similar to the mechanism by which VLA models bundle actions across multiple timesteps into a single chunk for generation.

Furthermore, the correspondence between reward prediction errors (analogous to the basal ganglia dopaminergic system) and RL's temporal difference (TD) error is not coincidental. Both systems use "the difference between what was expected and the actual outcome" as a learning signal to progressively improve behavior.

Practical Implications of This Analogy

This parallel is not merely an intellectual curiosity — it suggests future directions for VLA research:

• Humans' ability to flexibly extend their body schema (tool use) inspires cross-embodiment generalization research in VLA.
• The phenomenon of human motor memory being consolidated during sleep suggests the importance of offline RL and experience replay.
• Humans' ability to learn motor skills through observation alone (mirror neuron system) directly connects to learning latent actions from human video.

---

6.5 Self-improvement and Lifelong Learning

The ability of a VLA system to continue improving its performance after deployment in a real environment is one of the most important research directions from a practical standpoint.

Autonomous Data Collection

• SOAR (Fan et al., 2025): An autonomous data collection framework guided by foundation models (VLM, LLM). The model autonomously determines "which data is lacking" and autonomously collects data in those areas.
• Core idea: An embodied version of active learning — the model automatically explores and experiences situations of high uncertainty.

Online Self-improvement

• RoboCat [110] (Bousmalis et al., 2024): A pioneering system implementing a self-improvement loop. Successful trajectories generated by the model are added to the training data for iterative improvement.
• VLA-RL [68] (Tan et al., 2025): Continues to learn from environmental interaction after deployment through online RL.

The central challenge of self-improvement is self-reinforcement bias. When a model uses its own (imperfect) outputs as training data, existing errors or biases can be amplified. Quality filtering, diversity assurance mechanisms, and human-in-the-loop intervention are needed to prevent this.

The Core Challenge of Lifelong Learning: Catastrophic Forgetting

The most serious problem VLA systems face when adapting to new tasks or environments is catastrophic forgetting — the loss of previously learned capabilities when fine-tuned on new data.

Specific manifestations of catastrophic forgetting in VLA:

• Fully unfreezing the VL encoder and fine-tuning progressively erodes the rich visual-language understanding acquired during internet-scale pre-training.
• Overfitting to a specific environment degrades generalization ability in other environments.
• Specializing to a specific robot embodiment weakens cross-embodiment transfer capability.

Resolution strategies:

• Selective Unfreezing: Instead of updating all parameters, selectively fine-tunes only task-relevant layers. Parameter-efficient fine-tuning (PEFT) methods such as LoRA (Low-Rank Adaptation) are representative.
• ReVLA [111] (Shi et al., 2025): Introduces a reversible learning mechanism to reversibly preserve prior knowledge when learning new tasks.
• π0.5 [31]-KI: Selectively blocks gradient propagation to specific modules via gradient blocking, protecting pre-trained knowledge.

Cross-embodiment Generalization

Ultimately, VLA systems must be able to generalize to diverse embodiments rather than being tied to a specific robot. The skill of "grasping an object" learned on a 7-axis articulated robot must also work on a parallel gripper, a dexterous hand, and a mobile manipulator.

• HPT [96] (Wang et al., 2024): Heterogeneous Pretrained Transformer. An architecture that separates a shared latent space from embodiment-specific heads. The shared transformer processes task semantics, and a dedicated head for each robot morphology converts them into the corresponding action space.
• UniAct [112] (Ning et al., 2025): Defines a unified action space in 3D space, learning a universal action representation independent of robot morphology.
• BridgeVLA [113] (Li et al., 2025): A VLA model that acts as a bridge between different robot datasets, facilitating cross-dataset transfer.

The core challenge of cross-embodiment generalization is the heterogeneity of action spaces. A 7-DoF robot arm, a 12-DoF dexterous hand, and a 20+-DoF humanoid differ fundamentally in the dimensionality and meaning of their action spaces. To overcome this heterogeneity, latent action tokens (Type 6) or task-space representations are being used as key intermediaries.

---

6.6 Summary: Three Stages of Maturation in Learning Paradigms

Synthesizing the evolution of VLA learning paradigms reveals a three-stage maturation model strikingly similar to the developmental trajectory of LLM training:

VLA research is currently at the transition from Stage 2 to Stage 3. The results of RIPT-VLA [71] (4%→97% on a specific task per the original paper; LIBERO avg 74.7% per Jin et al. [9]) and SimpleVLA-RL [70] (17.3%→91.7%; original paper citation) suggest that this transition can deliver a paradigm-level leap, not merely incremental improvement. It is almost certain that RL post-training will join BC as a standard component of the VLA training pipeline going forward.

At the same time, as the analogy with human motor learning suggests, learning is not a single-stage problem but a continuous lifelong process. Post-deployment self-improvement, adaptation to new environments, and knowledge accumulation without catastrophic forgetting — implementing these lifelong learning capabilities is a long-term challenge for VLA research and an essential requirement on the path toward truly general-purpose robotic systems.

---

---

Enhancements for Chapters 5--6

E-1. Motivation Chain for Learning Paradigms

Understanding why VLA learning has evolved through three stages is best grasped as a chain of motivations, where each stage's limitations directly necessitate the next.

Stage 1 -- Pre-training: "See and understand the world." Motivation: A robot that has never seen a mug cannot be told to grasp one. Internet-scale image-text (and increasingly video) pre-training provides the broad semantic prior -- object recognition, spatial reasoning, physical commonsense -- that makes downstream robot learning sample-efficient.

Stage 2 -- Behavioral Cloning: "Imitate expert demonstrations." Motivation: World understanding alone does not produce motor commands. BC bridges perception and action by mapping observations to demonstrated actions via supervised learning. However, BC is bounded by demonstration quality and suffers from distribution shift and compounding errors.

Stage 3 -- RL Post-training: "Go beyond the demonstrations." Motivation: BC cannot discover better-than-demonstrated behaviors, handle multimodal action distributions gracefully, or incorporate safety and preference signals. RL post-training (PPO, GRPO, DPO) addresses all three, enabling the model to self-improve through trial-and-error or human feedback -- mirroring the LLM trajectory from SFT to RLHF.

Ongoing -- Lifelong Learning: "Never stop improving." Motivation: Deployment environments are non-stationary. The robot must adapt to new objects, tasks, and embodiments without catastrophic forgetting, closing the loop between experience and knowledge.

---

E-2. Comparison of Eight Action Token Types

Key takeaway: No single token type dominates. State-of-the-art VLA systems increasingly combine multiple types hierarchically -- e.g., reasoning tokens for planning + raw/latent tokens for execution.

---

E-3. Intuitive One-Liners

• Action tokenization is to VLA what vocabulary design is to an LLM: it determines what the model can say and how precisely it can say it.
• Behavioral Cloning is like learning to cook only by watching videos -- you can reproduce recipes you have seen, but you cannot improvise when an ingredient is missing.
• RL post-training is the deliberate practice session after the lecture: the robot tries, fails, and refines until it surpasses the instructor.
• Action chunking is the robotics equivalent of muscle memory: individual keystrokes become fluid words, individual timesteps become smooth trajectories.
• FAST tokenization treats action sequences the way MP3 treats audio: transform to the frequency domain, discard redundancy, and compress dramatically.
• Latent tokens are the Esperanto of robot actions: a universal language that lets different embodiments share the same motor concepts.
• Catastrophic forgetting is the price of plasticity -- a VLA that adapts too eagerly to a new task may forget everything it learned before.
• The sim-to-real gap is the uncanny valley of robotics data: simulation looks close enough to fool the eye, but not the physics.

---

E-4. Self-Check Questions

Q1. A VLA model trained with 256-bin raw action tokenization and cross-entropy loss consistently predicts the "average" action in bimodal scenarios (e.g., the robot hesitates instead of pushing left or right). (a) Explain the root cause of this failure mode. (b) Name two tokenization or decoding strategies that can mitigate it, and briefly explain why each works.

Q2. RIPT-VLA [71] reports a jump from 4% to 97% success rate after RL post-training (Su et al., 2025), yet Jin et al. [9] report a LIBERO average of 74.7% for the same model. Discuss at least two reasons why such a large discrepancy can arise when evaluating the same RL post-training method across different benchmarks or task suites.

Q3. Consider the neuroscientific analogy between basal ganglia chunking and VLA action chunking. (a) In what specific way does increasing the chunk size improve computational efficiency? (b) What practical risk does a large chunk size introduce, and how do systems like pi-0 [16] address it?

---

E-5. Open Research Questions

7. Efficiency — An Essential Challenge for Real-World Deployment

VLA (Vision-Language-Action) models are demonstrating remarkable performance on academic benchmarks, yet deploying them on physical robots in the field is an entirely different order of problem. These models must perform real-time inference with billions of parameters, on constrained hardware, while operating safely and economically. This chapter maps the full landscape of the efficiency problem, centering on the efficient VLA survey by Yu et al. [4] (2025).

---

7.1 Why Efficiency: The Gap Between Reality and Ideal

The resource demands of current VLA models are at an impractical level from the perspective of real-world deployment.

Scale of training costs:

• Training OpenVLA [15] required approximately 21,500 A100-GPU hours — equivalent to running a 64-GPU cluster continuously for about two weeks.
• Training π0 [16] used over 10,000 hours of robot trajectory data. Collecting data at this scale independently within a single institution is essentially infeasible.

The inference latency wall:

• RT-2-PaLI-X (55B) has an inference latency of 330–1000ms, corresponding to a control frequency of only 1–3Hz. This falls short even of the minimum frequency (5–10Hz) required for tabletop manipulation, let alone the 30Hz+ needed for dynamic tasks.
• Even the comparatively efficient OpenVLA [15] (7B) has a latency of 166ms (approximately 6Hz), making it unsuitable for tasks requiring rapid response.

Four key requirements for real-world deployment:

To close these gaps, research on efficient VLAs exploded from late 2024 through 2025. The research directions fall broadly into three axes: model efficiency, training efficiency, and data efficiency.

---

7.2 Model Efficiency: Making Inference Faster and Lighter

Model efficiency is a collective term for techniques that reduce latency and memory at the inference stage of an already-trained VLA. Five strategies exist: quantization, pruning, knowledge distillation, token optimization, and efficient architectures.

7.2.1 Quantization

Quantization is the most direct technique for reducing memory and computation by lowering the numerical precision of model weights (and activations).

• OpenVLA [15] 4-bit PTQ (Post-Training Quantization): Post-training quantization alone halved GPU memory usage with no observed performance degradation. This suggests that VLA model weights contain considerable numerical redundancy.
• SQIL [114] (Shang et al., 2024): Applying 4-bit salience-aware quantization achieved 2.5× inference speedup. The key is identifying weights critical for action prediction and selectively maintaining higher precision for those weights.
• BitVLA [33]: A study applying extreme 1-bit ternary quantization ({-1, 0, 1}), reporting 3.36× memory compression. It is notable that meaningful action generation remains possible when weights are represented with only three values.
• QAIL (Quantization-Aware Imitation Learning) [115] (Heo et al., 2025): Integrates quantization into the training phase to directly learn a model optimized for edge-device deployment.
• SQAP-VLA [116] (Li et al., 2025): Co-designs quantization and token pruning together, achieving a better efficiency-performance balance than applying each technique individually.

7.2.2 Pruning

Pruning is a technique for lightweighting a model by removing unnecessary components (layers, neurons, tokens, etc.). In VLAs, research is particularly active based on the observation that LLM backbone layers exhibit high redundancy.

Layer-level pruning:

• High cosine similarity is observed between the outputs of adjacent LLM layers, providing grounds for removing up to 50% of layers.
• DeeR-VLA [35]: Uses a dynamic early-exit strategy. It checks the consistency of action predictions at each layer and skips the remaining layers once consistency is confirmed. A major advantage is that no additional training is required.
• SmolVLA [32]: Takes an extremely simple approach, simply skipping L/2 layers of the LLM. It demonstrates that manipulation tasks can be performed with only half the layers.
• MoLe-VLA [117] (Qu et al., 2025): Uses a STAR router to dynamically select which layers to activate per input. Fewer layers are activated for easy tasks and more for complex ones, adaptively modulating computation.
• EfficientVLA [118] (Niu et al., 2025): A framework that simultaneously applies layer pruning and visual token pruning without training.
• FLOWER [119] (Cheng et al., 2025): In encoder-decoder VLMs, removes the entire decoder; in decoder-only architectures, removes the bottom 30% of layers.

Structured pruning:

• RLRC [120] (Zhao et al., 2025): Structured pruning based on Taylor importance scores, achieving up to 90% sparsity while maintaining meaningful performance.

7.2.3 Distillation

Distillation — transferring the knowledge of a large VLA into a smaller model — can achieve higher performance than training a small model from scratch.

• TinyVLA [34]: Distills from a large VLA into a sub-1.4B small model. Initializes with LoRA weights to improve distillation efficiency.
• CEED-VLA [121] (Wen et al., 2025): Combines consistency distillation with Jacobi parallel decoding. Parallelizes the serial bottleneck of autoregressive token generation, significantly improving inference speed.
• RPD (Robot Policy Distillation) [122] (Wang et al., 2025): Distills from a VLA into a small RL specialist policy. For specific tasks, a distilled specialist can be faster and more accurate than a general-purpose VLA.
• SP-VLA [123] (Shen et al., 2025): Uses action-aware scheduling to dynamically switch between a heavy VLA and a lightweight action generator. The large VLA is invoked only at moments requiring complex judgment; the lightweight generator handles straightforward execution phases.

7.2.4 Token Optimization

In VLAs, visual tokens account for the majority of the entire input sequence. A single image is converted into hundreds of patch tokens, and with video input this number can explode into the thousands. Reducing this is the core of token optimization.

Visual token compression:

• SmolVLA [32]: Compresses to 64 tokens per frame via pixel shuffle. Tokens that originally numbered in the hundreds are spatially rearranged and reduced to an extreme degree.
• FlashVLA [52] (Zhu et al., 2025): Removes low-importance visual tokens via ICS (Importance-based Compression and Selection) pruning.
• EfficientVLA: [118] Applies layer pruning and visual token pruning in an integrated fashion.

Visual token caching:

• VLA-Cache [124] (Gao et al., 2025), CronusVLA [125] (Lin et al., 2025): Exploit temporal coherence — tokens corresponding to a static background change very little between consecutive frames. By caching unchanging background tokens and updating only foreground tokens where change occurs, ~40-50% faster inference (per Zhang et al. [6]; original paper reports up to 2x+ acceleration) is achieved.
• The fundamental reason this approach is effective is that the majority of patch tokens are spatially redundant in robotic manipulation tasks. In tabletop environments with a fixed camera, more than 80% of the background is identical across frames.

7.2.5 Efficient Architectures

This is architectural-level innovation aimed at overcoming the fundamental limitations of existing Transformer structures (quadratic-complexity attention).

Linear-complexity architectures:

• SARA-RT [126] (Shridhar et al., 2024): Up-trains standard softmax attention into linear attention, reducing complexity from O(n²) to O(n).
• RoboMamba [127] (Liu et al., 2024): A VLA based on the Mamba SSM (Selective State Space Model), achieving more than 3× speedup at linear complexity. Its advantage over Transformers grows with longer sequences.

MoE (Mixture of Experts):

• GeRM [128] (Xu et al., 2025): Applies MoE to RL for quadruped robots, activating only a subset of expert parameters from the full parameter set.
• FedVLA [72] (Zhang et al., 2025): Implements efficient VLA in a federated learning setting using a dual-gating MoE.
• DriveMoE [49] (Huang et al., 2025): Leverages MoE in the autonomous driving domain to assign experts to various driving scenarios.

Parallel decoding:

• OpenVLA [15]-OFT: Uses bidirectional attention to generate multiple action tokens simultaneously.
• PD-VLA [129] (Chen et al., 2025): Parallelizes autoregressive decoding via Jacobi fixed-point iteration.
• Spec-VLA [130] (Wu et al., 2025): Applies speculative decoding to VLAs, achieving 1.42× speedup. A small draft model rapidly generates candidate tokens, which a large model then verifies.

7.2.6 Efficient Attention

Yu et al. [4] additionally identify Efficient Attention techniques as a distinct research direction. These methods optimize the Transformer attention mechanism itself, operating on an independent axis from model compression approaches such as quantization and pruning.

• KV-Efficient VLA: Compresses the KV cache using RNN-gated mechanisms, reducing the memory footprint of attention without discarding contextual information.
• Long-VLA [73]: Addresses long-horizon tasks via phase-aware input masking, selectively attending to task-relevant temporal segments rather than processing the entire observation history uniformly.
• RetoVLA [74]: Reuses register tokens across layers to avoid redundant computation in the attention mechanism.
• dVLA [65]: Applies prefix attention masking for diffusion-based VLAs, enabling efficient conditioning while preserving the generative quality of the diffusion action head.

These approaches are complementary to the compression techniques described above and can be combined with quantization, pruning, or token optimization for compounding efficiency gains.

---

7.3 Training Efficiency: Learning More with Fewer Resources

Alongside lightweighting the model itself, research into improving the efficiency of the training process is also active.

Parameter-Efficient Fine-Tuning (PEFT):

• PEFT methods including LoRA (Low-Rank Adaptation) train only 0.1–1% of total parameters while achieving performance comparable to full fine-tuning. This reduces GPU hours by approximately 70% and makes fine-tuning large VLAs feasible even on a single GPU.

Mixed training strategies:

• Curriculum learning: A strategy of progressively increasing task difficulty from easy to hard.
• Multi-stage training: π0 [16] uses a three-stage pipeline of (1) VLM pre-training → (2) robot data pre-training → (3) task-specific fine-tuning.

The innovation of FAST tokenization [20]:

• FAST (Fast Action Tokenization), proposed by Pertsch et al., applies DCT (Discrete Cosine Transform) + BPE (Byte Pair Encoding) to robot action sequences. This extremely compresses action sequences and accelerated pre-training by 5×. Using FAST tokens or latent actions instead of raw actions is a key trend in efficient action representation.

---

7.4 Data Efficiency: Learning More from Less Robot Data

The high cost of robot data collection is the most fundamental bottleneck in VLA research. Data efficiency research explores strategies that circumvent or alleviate this bottleneck.

Leveraging human video:

• Models such as EgoVLA [131] (Chen et al., 2025), Being-H0 [138] (Li et al., 2025), and RynnVLA-001 use first-person (ego-centric) human activity videos — abundant on the internet — as surrogate training data. Manipulation strategies are learned from human hand movements and transferred to robot actions. The core insight of this approach is that humans and robots perform similar manipulation tasks in the same physical world.

Simulation data:

• UniSim [101] (Yang et al., 2024), Genesis: Generate large-scale synthetic data from physics simulators.
• GraspVLA [81] (Qian et al., 2025): Generates billion-scale synthetic grasping data for use in pre-training.

Data augmentation:

• Language augmentation: Methods such as DIAL [90] paraphrase task instructions in diverse ways to improve the robustness of language understanding.
• Visual augmentation: Methods such as GenAug [87], CACTI [88], and ROSIE [89] use generative models to expand visual diversity.
• Trajectory augmentation: Methods such as DemoGen synthesize new trajectories from existing demonstration data.

Active data selection:

• AMF (Active Model Feedback): Prioritizes data with high information gain to maximize training efficiency.
• SWBT (Success Weighted by Trial): Includes failed attempts in training data, extracting useful signal even from failures.

Autonomous collection:

• SOAR (Luo et al., 2025): Robots autonomously collect data under the guidance of foundation models. This presents a pathway to continuously acquiring training data without human demonstrators.

---

7.5 Comparison of Key Lightweight Models

The table below compares representative VLA models by parameter scale, inference performance, and core techniques. It shows that over the course of a single year, compression progressed from 55B to 450M parameters, and from 1Hz to 120Hz.

---

7.6 Key Insights — Rediscovering the Efficiency-Performance Trade-off

The insights emerging from efficient VLA research go beyond mere technical optimization, calling for a reconsideration of VLA design philosophy itself.

1) The scale inversion phenomenon: The result that CLIP [27]-RT (~1B) outperforms OpenVLA [15] (7B) by 24% suggests that the simple application of scaling laws — "more parameters guarantee better performance" — may not hold in the robotics domain. Even a smaller model, when combined with appropriate representation learning and data-efficient fine-tuning, can surpass a much larger model.

2) Quantization is nearly a free lunch: The fact that 4-bit PTQ halves memory with no performance degradation means that considerable redundancy exists in current VLA weights. This provides a strong rationale for applying quantization by default at the deployment stage.

3) Hierarchical separation is uniquely well-suited to robotics: The asynchronous execution of a slow VLM (1–5Hz) + fast policy head (50Hz+) demonstrated by GR00T N1 [21] aligns naturally with the intrinsic structure of robot control. High-level semantic understanding does not need to be updated every frame, but low-level motor commands must be generated at high frequency. This "cognition slowly, action quickly" paradigm is also analogous to the structure of the human nervous system.

4) RL post-processing recovers compression losses: RIPT-VLA [71] demonstrated that RL post-processing can dramatically improve VLA performance. A BC/SFT (Behavioral Cloning / Supervised Fine-Tuning) baseline achieving only 4% was boosted to 97% through PPO post-processing (Su et al., 2025). Critically, the 4% to 97% result represents performance improvement of a BC/SFT baseline via RL — not the recovery of performance lost to quantization or pruning. Note that results vary by benchmark. This validates the feasibility of a pipeline: "lightweight model + RL post-processing."

5) Human video is a viable substitute for robot data: Research in the EgoVLA family shows that internet-scale human video can partially substitute for robot data. It is one of the most scalable pathways for circumventing the robot data collection bottleneck.

6) Dominant research trends: Efficient VLA research grew explosively from late 2024 through 2025. This reflects a shift in the research community from the strategy of "build it big first, shrink it later" toward "design it efficiently from the start."

---

7.7 Edge Deployment: System-Level Bottleneck Analysis

The 2026 Edge Embodied Foundation Models survey reframed VLA deployment as a systems engineering problem rather than a model compression problem. The "Deployment Gauntlet" proposed by this survey identifies seven coupled constraints that impede edge deployment: size, weight, power, memory traffic, compute latency, timing variability, and safety margins interact to form a compound problem that no single optimization can resolve.

The central finding is that the nature of the bottleneck differs by controller architecture:

• Autoregressive VLAs (RT-2, OpenVLA-class): Primarily constrained by memory bandwidth
• Diffusion-based controllers (pi-0-class): Primarily constrained by compute latency and sustained execution cost

This analysis suggests that architectures separating "fast control" from "slow semantic reasoning" (GR00T N1, π0.5 [31]) are advantageous for edge deployment as well. Efficient deployment requires system-level co-design that holistically considers memory architecture, scheduling strategies, communication protocols, and model design.

7.8 Complementary Efficiency Taxonomies

Guan et al. (2025) independently surveyed efficient VLAs and proposed a four-dimensional taxonomy: (1) model architecture, (2) perceptual feature extraction, (3) action generation mechanisms, and (4) training/inference strategies. While Yu et al. [4] focus on model compression and efficient design, Guan et al. treat the efficiency of perception and action generation as separate dimensions, making the two frameworks complementary.

Guan et al.'s [43] four-dimensional efficiency taxonomy provides an important complementary perspective for the efficiency analysis in this document. While Yu et al. [4] focus on model compression (quantization, pruning, distillation), Guan et al. analyze perceptual feature extraction efficiency (e.g., multi-resolution token pooling, selective attention) and action generation mechanism efficiency (e.g., Action Chunking optimization, parallel decoding) as independent dimensions. This perspective explains why recent models such as FlashVLA [52] and RetoVLA [74] pursue efficiency across the entire perception-action pipeline rather than simple model compression alone.

HyperVLA [135] (Park et al., 2026), proposed at ICLR 2026, uses hypernetworks to dynamically generate task-specific policies, accelerating inference. AutoQVLA [136] (Liu et al., 2026) achieves a 30% reduction in VRAM through improved quantization techniques. These works represent the cutting edge of the model efficiency techniques discussed in Section 7.2, demonstrating that quantization and architectural innovation remain active research directions.

---

---

8. Application Domains — The World VLA Is Making

VLA technology is spreading across a diverse range of robotic applications. Each domain has its own unique action space, safety requirements, and real-time constraints, and the manner in which VLAs are applied varies significantly as a result. This chapter surveys the major domains in which VLAs are currently being used, summarizing the current state of the art and the unique challenges of each area.

---

8.1 Tabletop Manipulation — The Mainstream Research Domain

Tabletop manipulation is the primary arena of VLA research. Over 70% of all VLA models are developed and evaluated in this domain.

Rapid improvement on benchmarks:

• LIBERO: Success rate rose from 76.5% to 98.1% within 16 months.
• CALVIN: Sequence length (number of consecutive tasks successfully completed) improved from 3.57 to 4.44.
• RLBench, Meta-World: Used as standard evaluation platforms for diverse manipulation tasks.

Current state and remaining challenges: Single-step manipulation tasks have nearly reached a solved state, with success rates above 98%. However, long-horizon tasks — tasks requiring multiple manipulation steps to be performed in sequence — remain a core bottleneck. The fundamental cause is the compounding error problem, in which errors accumulate across steps.

Specialized manipulation research:

• Bimanual manipulation: Models such as Bi-VLA (Xue et al., 2025) and ALOHA (Zhao et al., 2023) address the coordinated control of two arms. The action space doubles compared to a single arm, and synchronization between the two arms is the key challenge.
• Contact-rich manipulation: Models such as ForceVLA [78] (Lee et al., 2025) and TactileVLA (Kim et al., 2025) integrate force/tactile sensors into VLAs, detecting properties of objects — such as stiffness, weight, and slippage — that cannot be discerned through vision alone.
• Dexterous grasping: Models such as DexVLA (Wen et al., 2025) and DexVLG (Zhang et al., 2025) use VLAs to learn high-dimensional control of multi-finger hands. With degrees of freedom exceeding 20, the complexity of the action space increases sharply.

---

8.2 Humanoid Robots — The Challenge of Whole-Body Control

Applying VLAs to humanoid robots involves qualitatively different challenges from tabletop manipulation.

Fundamental difficulties:

• 30+ degrees of freedom: Control of the entire body — arms, legs, torso, and head — is required.
• Balance maintenance: The dynamic balance of bipedal locomotion demands fast responses on the order of milliseconds.
• Simultaneous locomotion and manipulation: Tasks such as picking up an object while walking require locomotion and manipulation to be performed simultaneously.
• Multi-contact-point management: Coordination is required across all contact points — feet, hands, and sometimes the torso — as they interact with the environment.

Key models:

• GR00T N1 [21] (NVIDIA): Positioned as a foundation model for humanoids. Achieves a high control frequency (~120Hz, motor output frequency; not reported in Yu et al. [4] Table 1; distinguish from model inference frequency) via its dual-system architecture, targeting general-purpose whole-body control.
• Humanoid-VLA [137] (Zhang et al., 2025): Performs pose estimation from human videos available online to secure diversity in motion. It uses human movement directly as reference data.
• Being-H0 [138] (Li et al., 2025): Uses ego-centric video as pre-training data to strengthen the ability to understand the environment from a first-person perspective.
• FP3 [139] (Chen et al., 2025): Strengthens spatial reasoning through 3D policy pre-training.

Key unsolved challenges: The simultaneous achievement of balance maintenance and precise manipulation remains one of the hardest challenges for current VLAs. Situations frequently arise in which the fast, reflexive control needed for balance conflicts with the deliberate, planned control needed for manipulation, and harmonizing these within a single unified model is the central challenge.

---

8.3 Autonomous Driving — Another Frontier for VLA

Autonomous driving is the second largest application domain for VLAs. According to the taxonomy of Jiang et al. [10], autonomous driving VLAs have undergone four stages of evolution.

The four stages of evolution:

Key models:

• EMMA [46] (Hwang et al., 2024): Waymo's end-to-end driving model using a Gemini backbone.
• ORION [47] (Wang et al., 2024): Combines a memory mechanism and CoT reasoning to leverage past driving experience.
• DriveMoE [49] (Huang et al., 2025): Uses a MoE structure to assign experts to various driving scenarios (highway, intersection, parking, etc.).
• AutoVLA [48] (Chen et al., 2025): Uses adaptive CoT — fast reasoning in simple situations, deeper reasoning in complex ones.

Driving vs. manipulation: key differences

Autonomous driving and robotic manipulation both share the VLA framework, but inherently involve very different challenges.

• SafeAuto [140] (Li et al., 2025) introduces a symbolic veto that blocks execution when the VLA's output violates a safety rule.
• LangCoop V2V [141] (Wei et al., 2025) addresses social interaction by sharing intent through vehicle-to-vehicle (V2V) natural language communication.

Benchmarks and remaining gaps: Benchmarks such as BDD100K, nuScenes, Bench2Drive, and Reason2Drive exist, but a key gap is the absence of a comprehensive "AI driving license" benchmark. There is as yet no standard that comprehensively evaluates diverse scenarios, safety judgment, and ethical dilemmas — analogous to a human driving license test.

---

8.4 Drones and Navigation

The application of VLAs is also expanding to aerial and ground mobile robots.

• CognitiveDrone [142] (Wang et al., 2025): A cognitive drone system that controls drones according to natural language instructions. It interprets and executes instructions such as "go around to the right of that red building."
• RaceVLA [143] (Zhao et al., 2025): Applies VLAs in the high-speed environment of drone racing. This is an extreme testbed in which millisecond-level reaction times and precise path tracking are simultaneously required.
• NaviLa [144] (Cheng et al., 2025), Uni-NaVid [145] (Zhang et al., 2024): Apply VLAs to indoor navigation of legged robots. They understand instructions such as "go to the kitchen and bring the red cup," performing path planning and obstacle avoidance.
• Mobility VLA [146] (Liu et al., 2025): A VLA for wheeled mobile robots that integrates autonomous navigation in indoor and outdoor environments with object interaction.

The common challenge in this domain is integrating real-time 3D navigation and dynamic obstacle avoidance within a single language-vision-action framework.

---

8.5 Medical and Surgical Robots

Medicine is a domain that simultaneously holds high potential for VLA application and the most stringent constraints.

Representative research:

• RoboNurse-VLA [147] (Li et al., 2024): Targets precision grasping in surgical environments. The precise grasping and transfer of surgical instruments is the core task.

Domain-specific constraints:

• Patient data privacy: Since the external transmission of medical data is legally restricted, on-premise inference is mandatory. VLA deployment strategies that rely on cloud APIs cannot be used in this domain.
• Small data problem: Data for specific surgical procedures or individual patients is inherently limited in quantity. The ability to effectively fine-tune a large pre-trained backbone on small data is critical.
• Safety-critical systems: Malfunction of a surgical robot directly affects patient life. Requirements for formal verification and safety assurance are more stringent than in any other domain.

These constraints mean that every dimension of efficiency (Chapter 7) — model lightweighting, on-device inference, and data efficiency — is especially urgent in the medical domain.

---

8.6 Agriculture and Industry

The potential of VLAs is also being explored in industrial applications of high practical value.

• Orchard apple harvesting (Zhang et al. [6]): A system in which a robot selectively harvests fruit according to natural language instructions ("harvest only ripe apples"). Simultaneous visual understanding in unstructured environments (branches, leaves, varying lighting) and gentle grasping are required.
• CIPHER (Park et al., 2025): A system that switches 3D printing inspection tasks via natural language instructions. It dynamically modifies the inspection procedure in response to instructions such as "inspect the surface quality of this part." This is a case of implementing process flexibility in industrial settings via VLA.
• ObjectVLA [148] (Chen et al., 2025): A VLA that can manipulate novel objects without prior demonstrations. It eliminates the cost of collecting demonstration data each time a new part or product is introduced on the factory floor.

The common requirement in the industrial domain is flexibility. In industrial settings where product types, task contents, and environmental conditions change frequently, the ability of VLAs to switch tasks on natural language instructions alone is of high practical value.

---

8.7 Interactive AR and GUI Agents

Beyond physical robots, the "action generation" capability of VLAs also extends to the autonomous manipulation of digital interfaces.

• ShowUI [149] (Lin et al., 2024): Implements a GUI (Graphical User Interface) agent using the VLA framework. It understands the visual content of a screen and generates actions such as clicking, scrolling, and typing in response to instructions like "open the settings menu and turn off Wi-Fi."
• Spatial grounding: Leverages the spatial understanding capability of VLAs to accurately place virtual objects in the physical world within AR (Augmented Reality) environments.
• Human-AI collaborative navigation: A scenario in which users and AI collaborate to navigate complex environments within augmented reality.

This domain demonstrates that the core components of VLA — visual understanding, language reasoning, and action generation — can serve as a powerful framework in domains beyond physical robots. It is an extension of the definition of "action" from physical motor commands to manipulation of digital interfaces.

---

8.8 Cross-Domain Comparative Summary

※ These frequency ranges are the author's synthesis of general domain requirements, not directly from individual surveys.

The pattern that stands out in this table is that VLA maturity is determined by an inverse relationship between data availability and safety requirements. Progress is fastest in tabletop manipulation, where data is abundant and safety constraints are low, and slowest in medicine, where data is scarce and safety is paramount. Closing this gap is a core challenge that must be addressed in the next stage of VLA research.

---

Motivation Chain: The Logic of Efficiency Research

Large VLAs are undeployable (RT-2 55B: 330-1000ms inference, hundreds of GB memory required)
--> Model reduction research begins (OpenVLA [15] 7B: 1/8 the size, performance retained)
--> 7B is still too heavy (16-24GB VRAM, 166ms latency)
--> Extreme lightweighting (SmolVLA [32] 450M, BitVLA [33] 1-bit quantization)
--> Concern over performance degradation from compression
--> Recovery via RL post-processing (lightweight model + RL = large-model-level performance)
--> Dynamic inference (DeeR-VLA [35]: shallow layers for easy inputs, deep layers for hard inputs)
--> Token optimization (FAST [20], VLA-Cache: reduce the amount of computation itself)

---

Efficiency Technique Comparison: Key Differentiators

---

Intuitive One-Liners: Efficiency and Applications

• Quantization: "Compressing a high-resolution photo to JPEG — the file shrinks dramatically, but it looks nearly the same to the eye."
• Pruning: "Trimming dead branches from a tree — the tree (model) becomes lighter and sways better in the wind (inference)."
• Distillation: "A master craftsman passing skills to an apprentice — the apprentice is smaller but preserves the core techniques."
• Token Caching (VLA-Cache): "Not redrawing the background every frame, just caching it — only the parts that change are recomputed."
• Dynamic Inference (DeeR-VLA [35]): "An exam strategy: answer easy questions quickly and move on, reserving deep thought only for hard problems."
• MoE (Mixture-of-Experts): "Not every doctor sees every patient — patients are routed to the relevant specialist. Each expert is small, but the collective capability is large."

---

Self-Check Questions: Sections 7-8

Q1: How does BitVLA's [33] 1-bit (ternary) quantization work, and why is performance preserved?

Answer: BitVLA [33] constrains model weights to three values: {-1, 0, +1} (ternary quantization). This replaces multiplication operations with additions and subtractions, drastically reducing memory and computation (3.36x compression). Performance is preserved because (1) quantization-aware training (QAT) compensates for quantization error during the learning process, and (2) VLA action outputs often do not require high numerical precision.

Q2: Identify three key domain differences between tabletop manipulation VLAs and autonomous driving VLAs.

Answer: (1) Safety level: Tabletop failures amount to object damage at most, whereas autonomous driving failures can cause loss of human life. (2) Control frequency: Tabletop tasks are adequately served at 5-50Hz, whereas autonomous driving mandates 30Hz+ real-time response as an automotive hardware standard. (3) Environmental diversity: Tabletop operates in a constrained workspace, whereas autonomous driving must handle infinitely diverse road conditions, weather, and traffic scenarios. These differences cause the two domains to evolve independently despite sharing the same underlying VLA architecture.

Q3: What does it mean that the LIBERO benchmark has reached "saturation," and what are the implications for VLA research?

Answer: Success rates on LIBERO-Object (99.8%) and LIBERO-Spatial (98.8%) have approached 100%, meaning this benchmark can no longer discriminate between model capabilities. This implies: (1) simple single-environment manipulation tasks have been effectively solved by VLAs, but (2) long-horizon composite tasks such as LIBERO-Long (96.6%) remain challenging, and (3) the community needs next-generation benchmarks that measure real-world generalization rather than in-distribution performance.

---

Open Research Questions: Sections 7-8

9. Datasets, Benchmarks, and Simulators

Progress in VLA research is not driven by architectural innovation alone. Large-scale datasets, reliable benchmarks, and realistic simulators must form a unified triad for research to advance. This chapter provides a systematic overview of the data infrastructure that underpins the VLA ecosystem.

---

9.1 Robot Learning Datasets

9.1.1 The Rise of Large-Scale Cross-Embodiment Datasets

In the early days of VLA research, it was common for individual research groups to construct small-scale datasets using their own robots and environments. MIME, RoboTurk, and RoboNet are representative products of this era. These datasets contained on the order of thousands to tens of thousands of episodes and focused on specific robot platforms and a limited range of tasks. However, unlocking the potential of large pretrained models required far larger and more diverse data.

The paradigm shift was spearheaded by the Open X-Embodiment [19] (OXE [19]) dataset. Assembled through a collaboration among 22 research institutions, it consolidates over one million episodes collected across 22 robot platforms into a single unified format — a dataset that deserves to be called the ImageNet of VLA research. Covering more than 527 skills, OXE [19] was the first large-scale demonstration of the viability of cross-embodiment learning. When RT-2-X was trained on OXE [19], it achieved more than a 50% performance improvement over models trained on a single dataset, dramatically illustrating the power of data diversity.

The following table summarizes the major robot learning datasets.

9.1.2 Strategic Use of Human Video Data

The cost of collecting robot data is overwhelmingly higher than that of internet text or images. One of the key strategies for circumventing this bottleneck is the use of human video data. Ego4D (3,670 hours), EPIC-Kitchens (100+ hours), and EgoDex (829 hours) capture manipulation activities performed by humans in everyday settings from an egocentric viewpoint, providing rich visual prior knowledge about "how to handle objects" without requiring the robot to collect it directly.

The case of GR-2, which achieved superior performance via human video pretraining followed by robot fine-tuning, and the case of HPT [96] (Wang et al., 2024), which improved cross-embodiment generalization by mixing human hand data with robot data, validate this strategy. However, the domain gap arising from morphological differences between humans and robots remains an outstanding challenge. EgoDex's dense 3D finger tracking data represents a concrete attempt to narrow this gap, providing human data in a form that can be directly applied to precise control of robotic hands.

9.1.3 Synthetic Data and Autonomous Collection

Another solution to the data bottleneck is synthetic data generation and autonomous collection. GraspVerse (source outside the 14 surveys) generated more than one billion synthetic grasp data samples, enabling large-scale pretraining in simulation. Autonomous collection pipelines such as SOAR (Luo et al., 2025; Self-supervised Autonomous Robot) demonstrate the possibility of expanding datasets without human supervision, through a self-supervised loop in which the robot explores, collects data, and labels it autonomously.

Of particular note is the fact that SmolVLA [32] (450M parameters), by focusing on data quality and curriculum, achieved performance competitive with far larger models. This suggests that data quality, diversity, and the design of the training curriculum may matter more than simple data scale — foreshadowing a paradigm shift from "more data" to "better data."

---

9.2 Simulation Benchmarks

9.2.1 Manipulation Benchmarks

Simulation benchmarks are a critical piece of infrastructure that enables systematic performance comparison of VLA models and rapid prototyping prior to real-world experiments. The following table summarizes the state of major benchmarks.

The LIBERO suite is one of the most widely used benchmarks in VLA research, offering four sub-benchmarks of increasing difficulty: Spatial (spatial relationship understanding), Object (object identification), Goal (goal achievement), and Long (long-horizon tasks). As of 2025, Spatial and Object have nearly reached saturation (98–99%). This indicates that VLA models have already achieved sufficient performance on short-horizon, simple manipulation tasks, suggesting that research focus should shift toward more complex long-horizon tasks.

CALVIN is a benchmark for evaluating multi-step sequential tasks, measuring the model's ability to execute up to five consecutive instructions. Performance is measured by Average Sequence Length, with DreamVLA [64] (Wen et al., 2025) recording the top score of 4.44. The dominant performance of world-model-based methods (VIP, DreamVLA [64], WorldVLA [76] (Chen et al., 2025)) on this benchmark corroborates the importance of future-prediction capabilities in long-horizon tasks.

Next-Generation Benchmarks (2025--2026)

New evaluation frameworks are emerging to address the saturation of existing benchmarks:

• RoboArena [150] (Li et al., 2025): A real-world-to-simulation automatic conversion framework that reproduces real-world tasks in simulation environments, enabling large-scale benchmarking.
• RoboCasa365 [151] (Nasiriany et al., 2025): A large-scale household environment benchmark comprising 365 tasks and over 2,000 kitchen scenes.
• WorldGym [152] (Zhang et al., 2025): A new paradigm that leverages action-conditioned world models as evaluation environments.
• WorldBench [41] (Hu et al., 2025): A unified evaluation platform for autonomous driving VLAs, integrating open-loop and closed-loop evaluation.

These benchmarks extend the evaluation paradigm beyond the saturation of LIBERO and CALVIN, toward measuring real-world generalization and domain diversity.

9.2.2 Autonomous Driving Benchmarks

Benchmarks for autonomous-driving VLAs have unique requirements distinct from those of manipulation. Safety, real-time performance, and adherence to social norms are the core evaluation axes.

Bench2Drive (Jia et al., 2024) is a closed-loop benchmark built on the CARLA simulator that evaluates agents' comprehensive driving capabilities across 220 routes and 44 scenarios. Models that score highly under open-loop evaluation frequently fail in the closed-loop setting, underscoring the necessity of closed-loop evaluation. Reason2Drive (Nie et al., 2024) is a new-paradigm benchmark that goes beyond simple path tracking to evaluate the reasoning process behind why a particular action was chosen.

---

9.3 The Simulator Ecosystem

The simulator ecosystem supporting VLA research has evolved diversely across domains.

Manipulation simulators:

• MuJoCo: The de facto standard for physics simulation. Its strengths are fast computation speed and accurate contact dynamics.
• SAPIEN: A simulator specialized for articulated object manipulation, supporting everyday-environment interactions such as opening drawers and operating faucets.
• RLBench: A benchmark-cum-simulator based on CoppeliaSim, providing more than 100 predefined tasks.
• AI2-THOR / Habitat: Simulators that combine indoor navigation with manipulation and serve as the primary platforms for embodied AI research.
• Isaac Gym (NVIDIA): Supports GPU-accelerated large-scale parallel simulation, enabling thousands of environments to run simultaneously and is optimized for RL training.

Autonomous driving simulators:

• CARLA: The leading open-source autonomous driving simulator. Supports diverse weather, traffic scenarios, and sensor modalities.
• nuPlan: A closed-loop planning benchmark and simulator based on the nuScenes dataset.

Next-generation general-purpose simulators:

• Genesis (source outside the 14 surveys): A GPU-accelerated physics engine that integrates various physics solvers and targets general-purpose robot simulation. Claims a 10–100× speedup over existing simulators.
• UniSim [101] (Yang et al., 2024): A "learned simulator" based on action-conditioned video diffusion, which learns environment dynamics directly from data without an explicit physics engine. This is an innovative approach that circumvents the realism limitations of traditional simulators.

Key gap: The greatest current limitation of the simulator ecosystem is the absence of a unified cross-embodiment, cross-task benchmark. Because each simulator uses its own task definitions, robot models, and evaluation protocols, directly comparing results reported across different simulators is practically impossible. This means that the unifying benchmark role that ImageNet played in computer vision has yet to be filled in robot learning.

---

9.4 Limitations of Evaluation Protocols and Directions for Improvement

9.4.1 The Reproducibility Crisis

Success-rate figures reported in VLA research are often misleading. In some studies, simply changing the random seed causes success rates to vary by more than 30%. This means that reported "state-of-the-art" numbers may not be statistically significant. The initial conditions of the environment, subtle variations in object placement, and non-determinism in the simulator's physics engine are among the causes of this variance.

9.4.2 The Simulation–Real-World Gap

The phenomenon of models that achieve high success rates in simulation failing in the real world remains pervasive. The main causes are inaccurate contact dynamics, limited visual realism, and unpredictable disturbances in the real world. The SIMPLER benchmark attempts to provide calibrated sim-to-real evaluation, but has yet to offer a fundamental solution.

9.4.3 Monolithic Evaluation Metrics

Most current benchmarks rely on a single metric — "success rate." However, this is insufficient when considering real-world deployment.

• Collision avoidance: Even if a task is completed, unnecessary collisions with the environment are dangerous.
• Failure recovery: The ability to return to a safe state upon failure is rarely reported.
• Energy efficiency: Completing the same task with less energy is practically important.
• Adversarial robustness: Resistance to intentional perturbations is essential in safety-critical applications.
• Inference latency: A model's inference speed determines whether real-time control is feasible, yet it is rarely reported systematically alongside success rates.

In autonomous driving, this problem is even more acute. There is no integrated "AI driving license" benchmark that simultaneously evaluates control safety and language fidelity, and proxy metrics such as open-loop L2 error have been repeatedly noted to correlate weakly with actual driving safety.

9.4.4 Proposed Improvements

To overcome these limitations, we propose a two-track evaluation framework.

(i) Simulation track: Standardized simulation evaluation sharing fixed seeds, data splits, and baseline models. Experimental configurations should be published in a fully reproducible form so that all research can be compared under identical conditions. Reporting mean and variance across a minimum of ten seeds should be mandatory.

(ii) Real-world community track: Real-world evaluation based on shared hardware protocols. The community should collectively define standardized robot platforms (e.g., Franka Emika, UR5), task definitions, and evaluation procedures, and each research group should report real-world performance using the same protocol.

---

10. Open Problems and Future Prospects

A cross-cutting analysis of ten major VLA survey papers identified eleven core challenges. These were addressed partially in individual surveys, but their full structure only emerges through cross-survey analysis.

---

10.1 The Data Bottleneck

The most fundamental constraint in VLA research is data. Even the largest existing dataset, OXE [19], contains only about 2.5 million episodes in its expanded version (the original v1 comprised 1M+ episodes) — a negligible amount compared to the training corpus of GPT-2 (WebText, billions of tokens). Moreover, unlike internet text, robot data costs tens of dollars per episode to collect and is bound to the unique morphology of each robot platform.

Directions for resolution:

• Simulation synthesis: Large-scale synthetic data generation such as GraspVerse (source outside the 14 surveys). Combined with domain randomization to facilitate sim-to-real transfer.
• Human video utilization: Extracting visual prior knowledge about manipulation from Ego4D, EPIC-Kitchens, etc.
• Autonomous collection (SOAR): A self-supervised pipeline in which the robot explores and collects data on its own.
• Active curation: Not all data is equal. Data is selectively collected by targeting the model's weaknesses.

Cross-survey insight: The case of SmolVLA [32] competing with 7B models despite having only 450M parameters suggests that data quality and diversity may matter more than data scale. This foreshadows a paradigm shift from "more data" to "better data."

---

10.2 The Generalization Wall

The generalization performance of VLA models changes dramatically with evaluation conditions.

• In-domain: 80–90% success rate under the same conditions as the training environment
• Cross-domain: Drops to 40–70% with novel objects or environments
• Zero-shot: Falls to 20–50% on entirely new tasks

Multiple approaches are underway to narrow this gap. HPT [96] (Wang et al., 2024; Heterogeneous Pretrained Transformers) attempts cross-embodiment generalization through pretraining across diverse embodiments; UniAct (Qian et al., 2025) attempts an embodiment-agnostic policy through a unified representation of the action space; and BridgeVLA (Li et al., 2025) attempts to connect web-scale visual knowledge with robot behavior.

Sim-to-real transfer remains an unsolved problem. Inaccurate physical contact, visual domain gaps, and the non-stationary nature of real-world environments are the primary barriers. The scaling-law study by GEN-0 (Team, 2025) provides early evidence that increasing model and data scale yields predictable improvements in generalization, but the limits of this law's validity remain unclear.

---

10.3 Real-Time Inference

The combination of large VLM backbones (7B–55B parameters) and diffusion-based action generation is powerful, but causes a critical latency problem for real-time control. Autonomous driving requires a minimum control frequency of 30Hz, and precise manipulation requires 50Hz or more, but even a single forward pass often makes it difficult to meet these requirements.

Solution strategies:

• Hierarchical asynchronous execution: The high-level VLM generates subgoals at a low frequency (1–5Hz) while a lightweight low-level policy performs actual control at a high frequency (50–100Hz). GR00T N1 [21], CogACT [23], and others adopt this approach.
• Token caching: Reusing key-value (KV) caches from previous inference passes to eliminate redundant computation.
• Quantization: Lowering model precision to FP16, INT8, or INT4 to increase inference speed. Cases in which performance degradation remains below 2% under 4-bit quantization have been reported.
• Pruning and distillation: Removing unnecessary parameters, or transferring knowledge from a large model to a small one.

The key concept is "intelligent sparsity." Rather than activating the entire model for every input, approaches that dynamically adjust the amount of computation according to input complexity are emerging.

---

10.4 Long-Horizon Tasks and Hierarchical Reasoning

Pure end-to-end models excel at single-action-level tasks but systematically fail on multi-step compositional tasks. Tasks such as "open the drawer, take out the cup, and place it on the shelf" require planning, subgoal setting, progress monitoring, and replanning upon failure — all of which are difficult for a single policy to handle.

Solution approaches:

• Hierarchical decomposition: π0.5 [31] explicitly separates a high-level VLM planner from a low-level action policy.
• Chain-of-Thought (CoT): CoT-VLA [55] inserts an explicit reasoning step before action generation, allowing the model to reason about why it selects a particular action.
• Skill library: Systems such as ReLEP (Park et al., 2025) learn reusable skill primitives and compose them to construct complex tasks.

Trends on the CALVIN benchmark validate this direction. World-model-based methods (DreamVLA [64]: 4.44, WorldVLA [76]: 4.38) show overwhelming performance over purely reactive policies, demonstrating that the ability to predict future states and use those predictions for planning is the key to success on long-horizon tasks.

---

10.5 Safety and Alignment

The safety challenges of VLAs are qualitatively different from those of purely software AI. Whereas hallucinations in an LLM result in incorrect text, hallucinations in a VLA can lead to physical collisions, damage, or even injury. The irreversibility of physical failures is the essential distinction.

Current attempts:

• SafeVLA [75] (Chen et al., 2025): The first attempt to explicitly integrate safety constraints into a VLA, combining safety-relevant training data with constraint-violation penalties.
• SafeAuto [140]: Implements a traffic-rule-based symbolic veto in autonomous driving. Neural network outputs are only executed once they pass a rule-based safety check — a dual-layer structure.

However, the absence of formal verification remains a serious challenge. Traditional control systems can guarantee safety through mathematical tools such as Lyapunov stability analysis and reachability analysis, but these verification methods have not been established for language-conditioned neural network policies.

This problem is particularly acute in autonomous driving. Whereas hallucinations in manipulation may amount to dropping an object, hallucinations in autonomous driving directly translate to traffic accidents. This "asymmetry of hallucination risk" means that autonomous-driving VLAs have fundamentally different safety requirements than manipulation VLAs.

---

10.6 Hallucination and Reasoning Stability

The problem of LLM-based planners generating physically impossible actions is a fundamental weakness of VLAs. Physically unreasonable plans such as "tilt the water glass 90 degrees and move it" arise when the LLM's common-sense reasoning diverges from physical reality.

SC-VLA [56] (Self-Correcting VLA) (Guo et al., 2025) introduced explicit failure detection and recovery reasoning mechanisms, reducing task failure rates by 35% (Zhang et al. [6]). It implements a feedback loop in which the model monitors the results of its own actions and generates an alternative action when an unexpected outcome is observed.

However, validating hallucinations in the open world is fundamentally difficult. To judge whether the model's output is "physically realizable" in situations not present in the training data, the model itself would need to function as an accurate physics simulator — a circular problem.

---

10.7 Multimodal Integration

Current VLA research exhibits a vision-centric bias. Most models use only RGB images as sensory input, largely ignoring the other senses that humans employ in manipulation: touch, force/torque, sound, temperature, and others.

• ForceVLA [78] (Lee et al., 2025): Integrates force/torque sensor data into the VLA, improving performance on delicate object manipulation.
• TactileVLA [153] (Kim et al., 2025): Uses tactile sensor inputs to perceive material properties (hardness, texture, etc.) that are difficult to judge from vision alone.
• OmniVTLA [154] (Wang et al., 2025): Proposes a unified architecture that simultaneously processes vision, touch, and language.

Humans automatically reweight modalities in high-uncertainty situations: relying more on touch when vision is insufficient, and focusing more on visual cues when the environment is noisy. This adaptive modality reweighting has yet to be systematically implemented in robots and represents an important future direction for multimodal VLAs.

---

10.8 Human–Robot Interaction

Current VLAs remain at the level of "pseudo-interaction." Unidirectional communication in which a human issues an instruction and the robot executes it is dominant; genuinely bidirectional, conversational collaboration is almost entirely unrealized.

True human–robot interaction requires the following:

• Adaptive dialogue: The robot asks clarifying questions about ambiguous instructions and adjusts its behavior based on human feedback.
• Preference learning: The robot learns human implicit preferences (speed, safety, aesthetic standards, etc.) through interaction.
• Human feedback loop: Continued improvement through human corrective feedback even after deployment.

Inspired by the success of RLHF (Reinforcement Learning from Human Feedback) in NLP, a new research direction of "RLHF for Robotics" is taking shape in this field.

---

10.9 Evaluation and Benchmarking

The evaluation limitations discussed in Section 9.4 merit re-examination at a more fundamental level as an open problem. The field of robot learning currently lacks a unified benchmark equivalent to ImageNet in computer vision or GLUE/SuperGLUE in NLP.

The consequences of this absence are serious. When Paper A reports 98% on LIBERO and Paper B reports 4.44 on CALVIN, there is no basis on which to judge which model is "better." Add to this the reproducibility issues caused by seed randomness, and quantitatively tracking genuine progress in VLA research becomes difficult in itself.

---

10.10 Ethics and Societal Impact

The real-world deployment of VLAs raises ethical and societal questions that go beyond technical challenges.

• Privacy: VLA robots operating in homes or workplaces continuously capture and interpret their environment. Clear guidelines for the collection, storage, and use of this data are needed.
• Job displacement: Advances in manipulation capabilities accelerate automation in logistics, manufacturing, and service industries, potentially causing structural changes in employment.
• Decision-making bias: Biases that VLM backbones have learned from internet data may manifest in physical actions. For example, biases related to race or gender could lead to discriminatory behavior in human–robot interaction.
• Regulatory framework: Outside of autonomous driving, regulatory frameworks for the deployment of VLA robots barely exist. Certification standards, liability attribution, and incident reporting systems need to be established urgently.

---

10.11 Cross-Survey Integrated Insights

The following ten insights, derived through cross-analysis of fourteen surveys, represent emergent patterns not explicitly visible in any individual survey.

Insight 1 — Evidence of Convergence

The fourteen surveys each use different taxonomies, but they are ultimately different projections of the same landscape. Architecture surveys classify VLAs along a "backbone–action head" axis; learning surveys along a "pretraining–fine-tuning" axis; application surveys along a "domain–task" axis. Yet across all these perspectives, "VLM Brain + Generative Action Head" is converging as the optimal point. This trend, which became clear from late 2024 through 2025, means that the "language model as action model" paradigm initiated by RT-2 [11] has now reached universal consensus.

Insight 2 — The Scale Inversion Phenomenon

The scaling-law intuition that "larger models yield better performance" does not necessarily hold for VLAs. Concrete evidence supports this.

• CLIP [27]-RT (1B) outperforms OpenVLA [15] (7B) on many tasks.
• SmolVLA [32] (450M) achieves performance competitive with 7B-class models on LIBERO.
• 3B-class models (CogACT [23], SpatialVLA [39]) achieve performance equal to or better than 7B models.

This suggests that data quality, tokenization strategy, and architectural design may matter more than parameter count. In VLAs, the scarcity of robot data can cause large models to overfit or waste unnecessary capacity.

Insight 3 — Tokenization Determines Control Bandwidth

The choice of action tokenization is not a mere implementation detail but a design decision that defines the fundamental capabilities of the system.

• Discrete bin tokenization: Simple to implement but limited in precision. Suitable for 1–5Hz control.
• Diffusion-based: Can represent continuous, multimodal distributions, but the iterative reverse-diffusion process degrades inference speed. Range: 5–20Hz.
• Flow matching: Achieves faster convergence than diffusion, enabling 20–50Hz.
• FAST tokeniz [20]ation: Simultaneously pursues the speed of discrete methods and the precision of continuous methods, enabling control frequencies up to 50–120Hz.

This perspective is a cross-cutting insight not explicitly addressed in any single survey. The choice of tokenization determines control frequency from 1Hz to 120Hz, and this fundamentally defines the range of tasks that can be performed. Slow control enables only coarse manipulation, while fast control makes high-difficulty tasks such as precision insertion, suturing, and playing musical instruments possible.

Insight 4 — Dual Systems Are Necessary, Not Optional

The limitations of pure end-to-end models on long-horizon tasks are repeatedly demonstrated. Daniel Kahneman's distinction between System 1 (fast, intuitive) and System 2 (slow, deliberate) has been validated as an engineering necessity in robotics.

GR00T N1 [21] adopts a dual-system architecture in which the high-level VLM (System 2) generates subgoals and the low-level diffusion policy (System 1) executes them. This yielded a 17% improvement in success rate and a 28% reduction in collision rate compared to a single-system approach — strong evidence that a theoretical distinction from cognitive science can be directly translated into an engineering design principle.

Insight 5 — RL Post-Training Is an Essential Complement to BC

VLAs trained on behavioral cloning (BC) alone have structural limitations. The distributional shift problem — in which performance degrades sharply when deviating from the distribution of the demonstration data — is a prime example. Reinforcement learning (RL) post-training has emerged as the key means of breaking through this limitation.

A dramatic example proves this: a case was reported in which a success rate that remained at 4% with SFT alone recovered to 97% after just 15 iterations of PPO (Proximal Policy Optimization). If BC teaches "what to do," RL trains the agent to learn "what is good." The combination of the two is not an option but a necessary pipeline step.

Insight 6 — The Efficiency–Performance Pareto Frontier Is Shifting

In 2025, the central competitive axis of VLA research has shifted from "absolute performance" to "compute efficiency." SmolVLA [32] (450M parameters) achieving performance comparable to what RT-2 [11] (55B parameters) achieved represents an efficiency revolution of more than 100×.

The "Intelligent Sparsity" paradigm is emerging. This is not simply about making models smaller, but about concentrating computation only where it is needed. LoRA-based efficient fine-tuning, Mixture-of-Experts (MoE) architectures, and early-exit mechanisms are the embodiment of this paradigm. "Performance per FLOP" is becoming the key metric, superseding simple scaling laws.

Insight 7 — Autonomous Driving VLAs Are on a Separate Evolutionary Path

Manipulation VLAs and autonomous-driving VLAs share the same "VLM + Action" framework, but in practice they follow quite different evolutionary paths. Autonomous driving has unique requirements compared to manipulation:

• Safety requirements: The consequences of failure are fatal, and social acceptance standards are far higher.
• Real-time requirements: A control frequency of 30Hz or more is absolutely non-negotiable.
• Social norms: Understanding and complying with social conventions such as traffic laws, yielding, and signal adherence is required.

It is regrettable that technology transfer between the two domains is insufficient. There are potential cross-pollination opportunities: precise control techniques from manipulation could contribute to fine steering in autonomous driving, and safety verification frameworks from autonomous driving could contribute to safety policies in manipulation.

Insight 8 — Human Motor Learning Theory as a Future Map for VLA Research

The mapping between Newell's motor learning theory and VLAs, proposed by Jin et al. [9], has the potential to function not merely as an analogy but as a systematic research framework. Newell's "freezing-freeing degrees of freedom" theory directly corresponds to a curriculum for hierarchical skill learning in VLAs that starts in a low-dimensional action space and gradually expands the degrees of freedom.

There are abundant unexplored territories:

• Robotic implementation of cerebellar models: The combination of a forward model and an inverse model from the human cerebellum could be translated into a combination of a world model and an inverse dynamics policy in VLAs.
• Contextual interference effects: The finding from human motor learning that random practice is superior to blocked practice for long-term retention could be applied to training curricula for VLAs.

Insight 9 — World Models Are the Key to Long-Horizon Tasks

The performance trends on the CALVIN benchmark deliver a clear message. Visual Interaction Prediction (VIP) methods show dominant performance, and approaches integrating world models — WorldVLA [76], DreamVLA [64], CoT-VLA [55] — dominate the top ranks.

World models implement the principle of "imagining before acting." Before executing a particular action, the system internally simulates how the world will change as a result, evaluates whether that outcome aligns with the goal, and only then executes the actual action. This fundamentally improves planning capability in long-horizon tasks and will be the key differentiator of the next generation of VLAs.

Insight 10 — The Rise of the "Defensive AI" Paradigm

Robustness is being elevated to a first-class design objective on par with performance. Even if a 98% success rate is achieved in the laboratory, a system that drops to 50% under unpredictable real-world perturbations cannot be deployed.

• BYOVLA (Build Your Own VLA): A modular architecture in which the robustness of each component can be independently verified and replaced.
• DreamVLA [64]: Imagination-based robustness improvement through a world model. Unexpected situations are internally simulated to prepare for them.
• SafeVLA [75]: Proactively blocks dangerous actions through explicit safety constraint integration.

In real-world deployment, robustness is not a "nice-to-have" property but a survival condition for the system. As this recognition spreads through the research community, the "Defensive AI" paradigm is taking shape.

---

10.12 The Generalization Gap Between Frontier and Open-Weight Models

As of 2026, the most prominent dividing line in the VLA field is the real-world generalization gap between proprietary frontier models (Gemini Robotics, π0.5 [31]) and open-weight research models. Although performance on simulation benchmarks (LIBERO, CALVIN) is converging between the two camps, a substantial gap persists in real-world zero-shot generalization. An analysis suggests that only the pi-family of models demonstrates competitive zero-shot behavior on the RoboArena leaderboard (Reuss, 2026).

Three factors are identified as causes of this gap: (1) Data quality and diversity gap -- the proprietary data of frontier labs exceeds public datasets in both quality and diversity; (2) Benchmark ceiling effect -- the saturation of simulation benchmarks masks genuine progress; and (3) Infrastructure scale gap -- the difference between laboratory-scale and industrial-scale training infrastructure.

At ICLR 2026, data quality curation and in-context learning were identified as the most underrepresented research directions, and these two directions may hold the key to closing the gap. This problem is closely linked to both the data bottleneck of Section 10.1 and the generalization wall of Section 10.2, and constitutes a central challenge for the open-source community.

---

10.13 Frontier Case Study: Two Breakthroughs from the π Series (2025.11 – 2026.03)

Sections 10.1–10.12 identified the core open problems and cross-survey insights for VLA research. How much concrete progress is actually being made on these challenges? In November 2025 and March 2026, Physical Intelligence (PI) published two papers in quick succession that provide the most tangible answer to this question. π^*_0.6 [157] directly attacks the "BC→RL transition" discussed in Section 6.3 and Insight 5, while π_0.6-MEM [158] confronts the "long-horizon tasks and memory absence" problem from Section 10.4. Both papers build on the π0.6 model (Gemma 3 4B VLM + 860M Action Expert) and break through two core VLA limitations—the performance ceiling of behavioral cloning and the absence of memory—at real-world scale.

10.13.1 π^*_0.6: A VLA That Learns from Experience

[157] · Physical Intelligence · 2025.11

π^*_0.6 [157] is a general-purpose RL post-training framework, built on a method called RECAP (RL with Experience and Corrections via Advantage-conditioned Policies), that enables VLA models to self-improve through real-world deployment experience. While the RL post-training methods surveyed in Section 6.3 (VLA-RL, RIPT-VLA, ConRFT, etc.) were primarily validated on simulation benchmarks, π^*_0.6 represents a qualitative turning point by being the first to successfully demonstrate end-to-end RL training of a large-scale VLA on real-world, long-horizon, complex manipulation tasks.

Core Technical Innovation: Advantage Conditioning. Unlike prior VLA RL methods that use policy gradient–based extraction such as PPO or GRPO, RECAP adopts advantage conditioning—a fundamentally different policy extraction mechanism. The key idea proceeds as follows:

(1) A distributional value function is trained separately. This value function uses a compact 670M-parameter VLM backbone and predicts the distribution of remaining steps to successful completion at each state (201 discrete bins). (2) The advantage value of each action is estimated from the value function and binarized into a text token "Advantage: positive/negative" appended to the VLA input. (3) The VLA is trained on all data (demonstrations + autonomous rollouts + human corrections) via advantage-conditioned supervised learning, but at inference time is always conditioned on "Advantage: positive" to extract the improved policy.

The decisive advantage of this approach is its compatibility with Flow Matching–based VLAs. PPO/GRPO require explicit computation of log-likelihoods, which Flow Matching models cannot directly provide and must approximate. Advantage conditioning completely bypasses this issue, achieving policy improvement through simple conditional supervised learning. In experiments, π^*_0.6 significantly outperformed AWR and PPO-based methods trained on the same data.

Unifying Three Data Sources. RECAP combines (1) demonstration data (for initial SFT), (2) autonomous rollouts (robot’s own attempts with success/failure labels), and (3) human corrections (human-gated DAgger—humans intervene during autonomous execution to correct mistakes) within a single framework. Human correction data always receives positive advantage, while the remaining data is assigned advantage based on the value function’s estimates.

Real-World Results. π^*_0.6 was validated on three complex tasks:

On the most challenging tasks, π^*_0.6 more than doubled throughput (successful completions per hour) and cut the failure rate roughly in half. Notably, a "targeted failure removal" experiment eliminated a specific failure mode (collar orientation error) to 97% success using only 600 autonomous trajectories over 2 iterations.

Significance in the Survey Context. π^*_0.6 is the most complete realization of the "BC→RL transition" discussed in Section 6.3. Where prior work demonstrated the feasibility of RL post-training on simulation benchmarks (LIBERO, etc.), π^*_0.6 is the first to prove the practicality of end-to-end RL for large-scale Flow Matching VLAs on real-world, long-horizon, complex tasks. This signals that the development trajectory from LLMs—GPT-3 (pre-training) → InstructGPT (SFT) → ChatGPT (RLHF)—is now materializing in VLA research.

10.13.2 π_0.6-MEM: Multi-Scale Embodied Memory for VLAs

[158] · Physical Intelligence · 2026.03

MEM (Multi-Scale Embodied Memory) endows VLAs with multi-modal, multi-timescale memory. Section 10.4 identified "long-horizon tasks and hierarchical reasoning" as a core open problem; MEM offers the most direct solution to this challenge.

Key Insight: Dual Memory Representation. When a robot executes a 15-minute task such as "clean the entire kitchen," two fundamentally different kinds of memory are required. (1) Short-term memory: recent visual information spanning a few seconds (when the arm occludes an object, or when recalling a failed grasp strategy). (2) Long-term memory: semantic events spanning several minutes (which ingredients have already been retrieved, which drawers have been opened). MEM’s core insight is that these two types of memory must be represented in different modalities.

Architectural Components:

(1) Short-term Video Memory (Video Encoder). The existing ViT is extended to process video input without adding any new learnable parameters. A space-time separable attention structure inserts causal temporal attention alongside spatial attention every 4th layer. Tokens from past timesteps are dropped in upper layers so that the total tokens passed to the VLA backbone remain identical to a single-frame VLA. This keeps inference latency within the 300ms real-time constraint even with 16-frame input (a naïve approach would exceed 4 seconds).

(2) Long-term Language Memory. A high-level policy compresses past semantic events into a natural-language summary (m_t) and incrementally updates it each step (m_t+1). The key is compression: "placed the light-green bowl, dark-blue bowl, and light-yellow bowl into the upper-right cabinet" → "placed three bowls into the upper-right cabinet," stripping unnecessary detail. This is critically important for reducing the train–inference distribution mismatch.

Core Design Principle: Initialization from Pre-trained VLM Weights. The video encoder is designed to guarantee exact equivalence with the original VLM at K=1 (single image) by setting the t=0 temporal position encoding to zero. This preserves the pre-trained VLM’s knowledge perfectly while adding memory capabilities. Experiments showed that introducing memory only during the post-training phase without pre-training caused marked degradation, confirming that diverse-data memory pre-training is essential.

Real-World Results.

Notably, the causal confusion problem repeatedly reported in prior work—where adding memory actually degrades performance—was not observed in π_0.6-MEM. This is attributed to the large-scale pre-training data mix encompassing diverse optimality levels, speeds, and control frequencies, which prevents spurious correlations.

Significance in the Survey Context. MEM directly addresses the reasoning paradigms of Section 4.2 (the brain module) and the core open problem of Section 10.4 (long-horizon tasks). While π0.5 approached long-horizon work through hierarchical separation ("VLM planner + VLA executor"), MEM tackles the same problem along the orthogonal dimension of memory. If hierarchical planning separates "what to do," memory remembers "what has been done." The two approaches are complementary rather than mutually exclusive, and their future combination is highly likely.

10.13.3 Integrated Significance of the Two Papers

Taken together, PI’s π series is simultaneously pushing two core frontiers of VLA research: "doing it better" (π^*_0.6) and "doing it longer" (MEM). π^*_0.6 elevates individual action quality beyond demonstration level, while MEM weaves individual actions into coherent task execution spanning tens of minutes. If the two approaches were combined in a single model, a robot that "self-improves at a 15-minute kitchen cleanup through trial and error" becomes feasible. This represents the most concrete advancement toward the "deployment readiness" paradigm put forth in this survey, demonstrating that the open problems identified above are transitioning from theoretical speculation to engineering challenges.

---

11. Conclusion

11.1 VLA — Realizing Unified Intelligence

Vision-Language-Action (VLA) models are the embodiment of a unified intelligence in which robots "see, understand, and act" in the world. By fusing visual perception, linguistic reasoning, and physical action within a single neural network, VLAs are fundamentally redefining the modular pipeline (perception–planning–control) of traditional robotics.

11.2 Three Years of History, 200 Models

Since RT-2 [11] first proposed the term "Vision-Language-Action Model" in 2023, more than 200 VLA models have emerged in just three years. This explosive growth is the result of three convergences: (1) the maturation of large language models, (2) advances in vision-language pretraining, and (3) the emergence of large-scale robot datasets. The Cambrian explosion of VLA research began in 2023–2024, when these three elements simultaneously reached a critical threshold.

11.3 Current Achievements and the Frontier

Short-horizon, single-domain manipulation tasks have nearly been solved. The figures of LIBERO-Spatial 98.8% and LIBERO-Object 99.8% show that the ability to perform defined tasks in defined environments has already approached human-level performance.

But the true frontier begins now.

• Long-horizon tasks: Planning and execution in multi-step compositional tasks
• Cross-domain generalization: Adaptation to environments and objects not encountered during training
• Real-world deployment: Stable operation in uncontrolled environments

These three challenges constitute the "last mile" of VLA research — and simultaneously the most difficult stretch.

11.4 The Efficiency Revolution

One of the most noteworthy trends in VLA research is the dramatic improvement in efficiency. The Pareto frontier is shifting: model size has been reduced by more than 100× from RT-2 [11]'s 55B parameters to SmolVLA [32]'s 450M parameters, while maintaining competitive performance. This is a decisive factor in accelerating the practical deployment of VLAs. Real-time inference on edge devices, cost-effective large-scale deployment, and energy efficiency are all beneficiaries of this efficiency revolution.

11.5 The Transition from BC to RL

A pipeline that begins with behavioral cloning (BC) and concludes with reinforcement learning (RL) post-training is becoming the standard for VLA training. The combination of the stable initialization provided by BC and the exploratory optimization provided by RL enables performance levels unattainable by either approach alone. The leap from 4% with SFT to 97% after 15 iterations of PPO strikingly illustrates the power of this combination.

11.6 Accelerating Domain-Specific Specialization

The general-purpose VLA framework is expanding into diverse domains.

• Autonomous driving: DriveVLM [91], DriveLM, and others implement VLAs specialized for road environments.
• Humanoids: GR00T N1 [21], HumanPlus, and others apply VLAs to whole-body control of humanoid robots.
• Medicine: VLA applications in surgical robots and rehabilitation assistance are being explored.

Each domain has its own unique safety requirements, control frequencies, and interaction patterns, and domain-specific design is accelerating accordingly.

11.7 Safety, Ethics, and Formal Verification — The Gatekeepers of Deployment

The final gatekeeper blocking real-world deployment of VLAs is not technical performance but safety and ethics. While efforts such as SafeVLA [75] and SafeAuto are underway, formal verification methods for language-conditioned neural network policies have yet to be established. This is a gateway that VLAs must pass through to move from the laboratory into everyday life, and it is an area that requires collaboration among regulatory bodies, industry, and academia.

Societal impacts such as privacy, job displacement, and decision-making bias must also be discussed in parallel with technological progress. It is too late to begin ethical deliberation only after a technology has been deployed in society.

11.8 Toward the Next Leap

The next leap in VLA research is expected to occur simultaneously along four axes.

First, world model integration. The ability to imagine outcomes before acting improves long-horizon tasks, safety, and generalization alike. The success of DreamVLA [64] and WorldVLA [76] validates this direction; in the next generation of VLAs, the world model will be a core module rather than an optional component (see Section 4.2.3 and the Large Model Embodied AI survey [44]).

Second, continual learning. Current VLAs are frozen after deployment, but a truly intelligent robot must continuously improve through experience. Continual learning — adapting to new tasks and environments without forgetting past learning (preventing catastrophic forgetting) — is the long-term vision for VLAs.

Third, general embodied intelligence. A universal policy in which a single model operates across diverse embodiments — robotic arms, humanoids, autonomous vehicles, drones — is the ultimate goal of VLA research. OXE [19] and HPT [96] have taken the first steps in this direction; scaling cross-embodiment generalization is the central challenge.

Fourth, human–robot co-evolution. As VLA robots become deeply integrated into human life, a co-evolution will begin in which humans and robots change each other. The feedback loop in which robots learn from human behavior and humans adjust their interaction patterns to match robot capabilities is the future that VLA research ultimately aspires to.

---

VLAs represent more than a technological advance; they are constructing an answer to the fundamental question of "whether machines can understand the physical world and act meaningfully within it." Three years after the naming of the field in 2023, this area has advanced at a remarkable pace, and that acceleration continues. The changes that the next three years will bring will surpass all that has come before.

---

Full VLA Taxonomy Tree

---

Motivation Chain: Deployment and Safety

Limitations of laboratory demonstrations (operate only in controlled environments; real-world deployment infeasible)
--> Efficiency research (lightweight models, quantization --> executable on edge devices)
--> Real-world deployment attempts (unexpected failure modes discovered)
--> Initiation of safety research (SafeVLA [75]: internalizing safety constraints into training)
--> Fundamental limits of safety (VLA hallucination --> possibility of physical accidents)
--> Need for formal verification emerges (still unresolved)

Limitations of single benchmarks (LIBERO saturation: simple tasks effectively solved)
--> Need for composite benchmarks (long-horizon, multi-step, real-world variation)
--> Simulation-to-real gap (sim-to-real gap persists)
--> Hybrid evaluation proposed (simulation + real-world + human evaluation)

---

Self-Check Questions: Sections 9--11

Q1: Explain the paradigm shift that the OXE dataset brought to the VLA field from the perspective of "data diversity."

Answer: Before OXE, each laboratory trained on small-scale data (thousands to tens of thousands of episodes) collected with its own robots. OXE unified over 1M episodes collected across 22 distinct robot embodiment types into a single format. The key finding was that data from diverse robots acts not as noise but as diversity, preventing overfitting to specific robots and environments and improving generalization. This is analogous to the phenomenon in NLP where multilingual training improves performance in each individual language.

Q2: Why is "hallucination" in VLAs fundamentally different from hallucination in LLMs?

Answer: LLM hallucination produces incorrect text, resulting in informational errors (stating things that are not factual). VLA hallucination generates erroneous actions that are executed in the physical world, meaning the consequences can include physical accidents (collisions, damage, injury). The difference is between "stating a nonexistent historical fact" and "swinging a robot arm to grasp an object that does not exist." For this reason, safety challenges in VLAs are fundamentally more severe than in LLMs, and the need for formal verification is correspondingly greater.

Q3: What is the greatest limitation of the current VLA benchmark ecosystem?

Answer: (1) Absence of a unified cross-benchmark: There is no standard benchmark equivalent to ImageNet or SuperGLUE, making it impossible to directly compare results across different simulators (LIBERO, CALVIN, RLBench). (2) Simulation-to-real gap: High performance in simulation does not guarantee real-world performance. (3) Saturation problem: Simple task benchmarks have already reached 99%, losing discriminative power. (4) Absence of long-horizon and unstructured task evaluation: Benchmarks for composite tasks lasting 30+ minutes or measuring the ability to cope with unexpected situations remain scarce.

---

Open Research Questions: Sections 9--11

References

1. [1] Ma, Q. et al. (2024). A Survey on Vision-Language-Action Models for Embodied AI. arXiv:2405.14093. [arXiv]
2. [2] Kawaharazuka, K. et al. (2025). Real-World Robot Applications of Foundation Models: A Review. arXiv:2402.05741. [arXiv]
3. [3] Zhong, Z. et al. (2025). Pure Vision Language Action (VLA) Models: A Comprehensive Survey. arXiv:2509.19012. [arXiv]
4. [4] Yu, Z. et al. (2025). A Survey on Efficient Vision-Language-Action Models. arXiv:2510.24795. [arXiv]
5. [5] Liu, N. & Shao, R. et al. (2025). Large VLM-based VLA Models for Robotic Manipulation: A Survey. arXiv:2508.13073. [arXiv]
6. [6] Zhang, Y. et al. (2025). VLA Models: Concepts, Progress, Applications and Challenges. arXiv:2505.04769. [arXiv]
7. [7] Chen, Y. et al. (2025). A Survey on VLA Models: An Action Tokenization Perspective. arXiv:2507.01925. [arXiv]
8. [8] Xu, C. et al. (2025). An Anatomy of Vision-Language-Action Models. arXiv:2512.11362. [arXiv]
9. [9] Jin, A. et al. (2025). Parallels Between VLA Model Post-Training and Human Motor Learning. arXiv:2506.20966. [arXiv]
10. [10] Jiang, H. et al. (2025). A Survey on VLA Models for Autonomous Driving. arXiv:2506.24044. [arXiv]
11. [11] Brohan, A. et al. (2023). RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. arXiv:2307.15818. [arXiv]
12. [12] Brohan, A. et al. (2022). RT-1: Robotics Transformer for Real-World Control at Scale. arXiv:2212.06817. [arXiv]
13. [13] Reed, S. et al. (2022). A Generalist Agent (Gato). arXiv:2205.06175. [arXiv]
14. [14] Ahn, M. et al. (2022). Do As I Can, Not As I Say: Grounding Language in Robotic Affordances (SayCan). arXiv:2204.01691. [arXiv]
15. [15] Kim, M. et al. (2024). OpenVLA: An Open-Source Vision-Language-Action Model. arXiv:2406.09246. [arXiv]
16. [16] Black, K. et al. (2024). pi0: A Vision-Language-Action Flow Model for General Robot Control. arXiv:2410.24164. [arXiv]
17. [17] Chi, C. et al. (2023). Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. arXiv:2303.04137. [arXiv]
18. [18] Driess, D. et al. (2023). PaLM-E: An Embodied Multimodal Language Model. arXiv:2303.03378. [arXiv]
19. [19] Open X-Embodiment Collaboration (2023). Open X-Embodiment: Robotic Learning Datasets and RT-X Models. arXiv:2310.08864. [arXiv]
20. [20] Pertsch, K. et al. (2025). Fast Tokenizer for VLA (pi0-FAST). arXiv:2501.09747. [arXiv]
21. [21] Bjorck, J. et al. (2025). GR00T N1: An Open Foundation Model for Generalist Humanoid Robots. arXiv:2503.14734. [arXiv]
22. [22] Huang, W. et al. (2023). Inner Monologue: Embodied Reasoning through Planning with Language Models. arXiv:2207.05608. [arXiv]
23. [23] Liu, H. et al. (2024). CogACT: A Foundational VLA Model with Cognitive-Inspired Action Chunking Transformer. arXiv:2411.19650. [arXiv]
24. [24] Liu, H. et al. (2024). RDT-1B: A Diffusion Foundation Model for Bimanual Manipulation. arXiv:2410.07864. [arXiv]
25. [25] Team, Octo Model et al. (2024). Octo: An Open-Source Generalist Robot Policy. arXiv:2405.12213. [arXiv]
26. [26] Shridhar, M. et al. (2021). CLIPort: What and Where Pathways for Robotic Manipulation. arXiv:2109.12098. [arXiv]
27. [27] Radford, A. et al. (2021). Learning Transferable Visual Models From Natural Language Supervision (CLIP). arXiv:2103.00020. [arXiv]
28. [28] Dosovitskiy, A. et al. (2020). An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ViT). arXiv:2010.11929. [arXiv]
29. [29] Brown, T. et al. (2020). Language Models are Few-Shot Learners (GPT-3). arXiv:2005.14165. [arXiv]
30. [30] Oquab, M. et al. (2023). DINOv2: Learning Robust Visual Features without Supervision. arXiv:2304.07193. [arXiv]
31. [31] Physical Intelligence (2025). pi0.5: a Vision-Language-Action Model with Open-World Generalization. arXiv:2503.01222. [arXiv]
32. [32] Pertsch, K. et al. (2025). SmolVLA: A Small Vision-Language-Action Model for Efficient Robot Learning. arXiv:2506.01844. [arXiv]
33. [33] Ma, Y. et al. (2025). BitVLA: 1-bit Vision-Language-Action Models. arXiv:2505.07256. [arXiv]
34. [34] Wu, J. et al. (2025). TinyVLA: Towards Fast and Data-Efficient VLA. arXiv:2409.12514. [arXiv]
35. [35] Yue, W. et al. (2024). DeeR-VLA: Dynamic Inference of Multimodal LLMs for Efficient Robot Execution. arXiv:2411.02359. [arXiv]
36. [36] Liang, J. et al. (2023). Code as Policies: Language Model Programs for Embodied Control. arXiv:2209.07753. [arXiv]
37. [37] Zhen, H. et al. (2024). 3D-VLA: A 3D Vision-Language-Action Generative World Model. arXiv:2403.09631. [arXiv]
38. [38] Huang, W. et al. (2022). Language Models as Zero-Shot Planners (SayTap). arXiv:2201.07207. [arXiv]
39. [39] Xu, Z. et al. (2024). SpatialVLA: Exploring Spatial Representations for VLA Models. arXiv:2501.15830. [arXiv]
40. [40] Wen, B. et al. (2024). TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for VLA. arXiv:2412.10345. [arXiv]
41. [41] Hu, T. et al. (2025). Vision-Language-Action Models for Autonomous Driving: Past, Present, and Future. arXiv:2512.16760. [arXiv]
42. [42] Edge Survey (2026). Embodied Foundation Models at the Edge: A Survey of Deployment Constraints and Mitigation Strategies. arXiv:2603.16952. [arXiv]
43. [43] Guan, W. et al. (2025). Efficient Vision-Language-Action Models for Embodied Manipulation: A Systematic Survey. arXiv:2510.17111. [arXiv]
44. [44] Large Model Embodied AI (2025). Large Model Empowered Embodied AI: A Survey on Decision-Making and Embodied Learning. arXiv:2508.10399. [arXiv]
45. [45] Jiang, Y. et al. (2022). VIMA: General Robot Manipulation with Multimodal Prompts. arXiv:2210.03094. [arXiv]
46. [46] Hwang, J. et al. (2024). EMMA: End-to-End Multimodal Model for Autonomous Driving. arXiv:2410.23262. [arXiv]
47. [47] Fu, H. et al. (2025). ORION: A Holistic End-to-End Autonomous Driving Framework. arXiv:2503.19755. [arXiv]
48. [48] Zhou, X. et al. (2025). AutoVLA: Autonomous Driving with Adaptive Reasoning and RL Fine-Tuning. arXiv:2506.13757. [arXiv]
49. [49] Yang, Z. et al. (2025). DriveMoE: Mixture-of-Experts for End-to-End Autonomous Driving. arXiv:2505.16278. [arXiv]
50. [50] Doshi, R. et al. (2024). CrossFormer: Scaling Cross-Embodied Learning. arXiv:2408.11812. [arXiv]
51. [51] Wu, H. et al. (2023). GR-1: Unleashing Large-Scale Video Generative Pre-training for Visual Robot Manipulation. arXiv:2312.13139. [arXiv]
52. [52] Tan, Z. et al. (2025). FlashVLA: Token-Aware Compression and Action Reuse for Efficient VLA Inference. arXiv:2505.21200. [arXiv]
53. [53] Zheng, K. et al. (2025). X-VLA: Cross-Embodiment Vision-Language-Action Model. arXiv:2510.10274. [arXiv]
54. [54] Du, Y. et al. (2025). HiMoE-VLA: Hierarchical Mixture-of-Experts for Generalist VLA Policies. arXiv:2512.05693. [arXiv]
55. [55] Zhao, Y. et al. (2025). CoT-VLA: Visual Chain-of-Thought Reasoning for VLA Models. arXiv:2503.22020. [arXiv]
56. [56] Li, X. et al. (2024). SC-VLA: A Self-Correcting VLA Model for Fast and Slow System Manipulation. arXiv:2405.17418. [arXiv]
57. [57] Huang, W. et al. (2023). VoxPoser: Composable 3D Value Maps for Robotic Manipulation. arXiv:2307.05973. [arXiv]
58. [58] Huang, S. et al. (2024). A3VLM: Actionable Articulation-Aware Vision Language Model. arXiv:2406.07549. [arXiv]
59. [59] Black, K. et al. (2023). SuSIE: Zero-Shot Robotic Manipulation with Pretrained Image-Editing Diffusion Models. arXiv:2310.10639. [arXiv]
60. [60] Lee, S. et al. (2024). VQ-BeT: Behavior Generation with Latent Actions. arXiv:2403.03181. [arXiv]
61. [61] Ye, D. et al. (2024). LAPA: Latent Action Pretraining from Videos. arXiv:2410.11758. [arXiv]
62. [62] Gu, Y. et al. (2023). RT-Trajectory: Robotic Task Generalization via Hindsight Trajectory Sketches. arXiv:2311.01977. [arXiv]
63. [63] Williams, R. et al. (2025). Lite VLA: Efficient VLA Control on CPU-Bound Edge Robots. arXiv:2511.05642. [arXiv]
64. [64] Zhang, H. et al. (2025). DreamVLA: A VLA Model Dreamed with Comprehensive World Knowledge. arXiv:2507.04447. [arXiv]
65. [65] Wen, C. et al. (2025). dVLA: Diffusion VLA with Multimodal Chain-of-Thought. arXiv:2509.25681. [arXiv]
66. [66] Chen, Z. et al. (2025). TGRPO: Fine-tuning VLA via Trajectory-wise Group Relative Policy Optimization. arXiv:2506.08440. [arXiv]
67. [67] Huang, J. et al. (2025). ThinkAct: VLA Reasoning via Reinforced Visual Latent Planning. arXiv:2507.16815. [arXiv]
68. [68] Lu, Y. et al. (2025). VLA-RL: Towards Masterful Robotic Manipulation with Scalable RL. arXiv:2505.18719. [arXiv]
69. [69] Chen, R. et al. (2025). ConRFT: A Reinforced Fine-tuning Method for VLA via Consistency Policy. arXiv:2502.05450. [arXiv]
70. [70] Li, Q. et al. (2025). SimpleVLA-RL: Scaling VLA Training via Reinforcement Learning. arXiv:2509.09674. [arXiv]
71. [71] Tan, W. et al. (2025). RIPT-VLA: Interactive Post-Training for VLA Models. arXiv:2505.17016. [arXiv]
72. [72] Miao, L. et al. (2025). FedVLA: Federated VLA Learning with Dual Gating MoE. arXiv:2508.02190. [arXiv]
73. [73] Fan, Y. et al. (2025). Long-VLA: Unleashing Long-Horizon Capability of VLA for Robot Manipulation. arXiv:2508.19958. [arXiv]
74. [74] Koo, J. et al. (2025). RetoVLA: Reusing Register Tokens for Spatial Reasoning in VLA. arXiv:2509.21243. [arXiv]
75. [75] Zhang, S. et al. (2025). SafeVLA: Towards Safety Alignment of VLA via Constrained Learning. arXiv:2503.03480. [arXiv]
76. [76] Cen, J. et al. (2025). WorldVLA: Towards Autoregressive Action World Model. arXiv:2506.21539. [arXiv]
77. [77] Li, Z. et al. (2025). PointVLA: Injecting the 3D World into VLA Models. arXiv:2503.07511. [arXiv]
78. [78] Yu, F. et al. (2025). ForceVLA: Enhancing VLA with Force-aware MoE for Contact-rich Manipulation. arXiv:2505.22159. [arXiv]
79. [79] Liu, J. et al. (2025). HybridVLA: Collaborative Diffusion and Autoregression in a Unified VLA Model. arXiv:2503.10631. [arXiv]
80. [80] Bu, Z. et al. (2025). UniVLA: Learning to Act Anywhere with Task-centric Latent Actions. arXiv:2505.06111. [arXiv]
81. [81] Deng, Y. et al. (2025). GraspVLA: Grasping Foundation Model Pre-trained on Billion-scale Synthetic Data. arXiv:2505.03233. [arXiv]
82. [83] Tian, R. et al. (2023). RAPL: What Matters to You? Visual Representation Alignment for Robot Learning. arXiv:2310.07932. [arXiv]
83. [84] Xia, Z. et al. (2025). HAPO: Human-assisted Robotic Policy Refinement via Action Preference Optimization. arXiv:2506.07127. [arXiv]
84. [85] Patel, D. et al. (2025). IKER: Real-to-Sim-to-Real with VLM-Generated Iterative Keypoint Rewards. arXiv:2502.08643. [arXiv]
85. [86] Xu, J. et al. (2025). KV-Efficient VLA: Speed up VLMs with RNN-Gated Chunked KV Cache. arXiv:2509.21354. [arXiv]
86. [87] Chen, X. et al. (2023). GenAug: Retargeting Behaviors to Unseen Situations via Generative Augmentation. arXiv:2302.06671. [arXiv]
87. [88] Mandi, Z. et al. (2022). CACTI: A Framework for Scalable Multi-Task Multi-Scene Visual Imitation Learning. arXiv:2212.05711. [arXiv]
88. [89] Yu, T. et al. (2023). ROSIE: Scaling Robot Learning with Semantically Imagined Experience. arXiv:2302.11550. [arXiv]
89. [90] Xiao, T. et al. (2022). DIAL: Robotic Skill Acquisition via Instruction Augmentation with VLMs. arXiv:2211.11736. [arXiv]
90. [91] Tian, X. et al. (2024). DriveVLM: The Convergence of Autonomous Driving and Large VLMs. arXiv:2402.12289. [arXiv]
91. [92] Zawalski, K. et al. (2024). ECoT: Robotic Control via Embodied Chain-of-Thought Reasoning. arXiv:2407.08693. [arXiv]
92. [93] Du, Y. et al. (2023). UniPi: Learning Universal Policies via Text-Guided Video Generation. arXiv:2302.00111. [arXiv]
93. [94] Nematollahi, I. et al. (2025). LUMOS: Language-Conditioned Imitation Learning with World Models. arXiv:2503.10370. [arXiv]
94. [95] Chi, B. et al. (2025). MinD: Learning A Dual-System World Model for Real-Time Planning. arXiv:2506.18897. [arXiv]
95. [96] Wang, L. et al. (2024). HPT: Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers. arXiv:2409.20537. [arXiv]
96. [97] Ross, S. et al. (2011). DAgger: A Reduction of Imitation Learning to No-Regret Online Learning. arXiv:1011.0686. [arXiv]
97. [98] Hancock, W. et al. (2025). Actions as Language: Fine-Tuning VLMs into VLAs Without Catastrophic Forgetting. arXiv:2509.22195. [arXiv]
98. [99] GraspVerse (2025). Large-scale Synthetic Grasp Data Generation. (14개 서베이 외 출처)
99. [100] Cheang, C. et al. (2024). GR-2: A Generative Video-Language-Action Model with Web-Scale Knowledge for Robot Manipulation. arXiv:2410.06158. [arXiv]
100. [101] Yang, S. et al. (2023). UniSim: Learning Interactive Real-World Simulators. arXiv:2310.06114. [arXiv]
101. [102] Singh, I. et al. (2023). ProgPrompt: Generating Situated Robot Task Plans using Large Language Models. arXiv:2209.11302. [arXiv]
102. [103] Wang, G. et al. (2023). Voyager: An Open-Ended Embodied Agent with Large Language Models. arXiv:2305.16291. [arXiv]
103. [104] Vemprala, S. et al. (2024). ChatGPT for Robotics: Design Principles and Model Abilities. arXiv:2306.17582. [arXiv]
104. [105] Nasiriany, S. et al. (2024). RT-Affordance: Affordances are Versatile Intermediate Representations for Robot Learning. arXiv:2411.02704. [arXiv]
105. [106] Xu, Z. et al. (2025). A0: An Autonomous Agent with Adaptive Action Generation. arXiv:2504.12636. [arXiv]
106. [107] Wang, H. et al. (2025). VQ-VLA: Vector Quantized Vision-Language-Action Model. arXiv:2507.01016. [arXiv]
107. [108] Liu, B. et al. (2025). Embodied-R1: Incentivizing Reasoning in Embodied VLA Models. arXiv:2508.13998. [arXiv]
108. [109] Wang, Z. et al. (2025). GRAPE: Generalizing Robot Policy via Preference Alignment. arXiv:2411.19309. [arXiv]
109. [110] Bousmalis, K. et al. (2024). RoboCat: A Self-Improving Generalist Agent for Robotic Manipulation. arXiv:2306.11706. [arXiv]
110. [111] Shi, L. et al. (2025). ReVLA: Reverting Visual Domain from LLM to VLA. arXiv:2409.15250. [arXiv]
111. [112] Zheng, Z. et al. (2025). UniAct: Universal Action Representation for Robotic Learning. arXiv:2501.10105. [arXiv]
112. [113] Li, J. et al. (2025). BridgeVLA: Bridging the Gap Between VLA and Low-Level Robot Control. arXiv:2506.07961. [arXiv]
113. [114] Park, D. et al. (2025). SQIL: Sub-4-bit Quantization of Large VLAs via Self-play Fine-tuning. arXiv:2505.15304. [arXiv]
114. [115] Heo, J. et al. (2025). QAIL: Quantization-Aware Imitation Learning for Resource-Efficient VLA. arXiv:2412.01034. [arXiv]
115. [116] Li, S. et al. (2025). SQAP-VLA: Stochastic Quantization with Adaptive Precision for VLA. arXiv:2509.09090. [arXiv]
116. [117] Qu, L. et al. (2025). MoLe-VLA: Mixture of Lightweight Experts for VLA. arXiv:2503.20384. [arXiv]
117. [118] Niu, X. et al. (2025). EfficientVLA: An Efficient Vision-Language-Action Model. arXiv:2506.10100. [arXiv]
118. [119] Cheng, Y. et al. (2025). FLOWER: Flow-based World Model for Efficient Robot Learning. arXiv:2509.04996. [arXiv]
119. [120] Zhao, Y. et al. (2025). RLRC: Reinforcement Learning with Reasoning Consistency for VLA. arXiv:2506.17639. [arXiv]
120. [121] Wen, Z. et al. (2025). CEED-VLA: Confidence-Enhanced Early-Exit Decoding for VLA. arXiv:2506.13725. [arXiv]
121. [122] Julg, M. et al. (2025). RPD: Robot Policy Distillation from Vision-Language-Action Models. arXiv:2503.05833. [arXiv]
122. [123] Shen, W. et al. (2025). SP-VLA: Spatial-aware Parallel Decoding VLA. arXiv:2506.12723. [arXiv]
123. [124] Xu, Y. et al. (2025). VLA-Cache: Accelerating VLA Inference via KV Cache Compression. arXiv:2502.02175. [arXiv]
124. [125] Lin, X. et al. (2025). CronusVLA: Efficient VLA with Temporal Cronus Attention. arXiv:2506.19816. [arXiv]
125. [126] Shridhar, M. et al. (2024). SARA-RT: Scaling Up Robot Action with Linear Attention. arXiv:2312.01990. [arXiv]
126. [127] Liu, Y. et al. (2024). RoboMamba: Efficient Vision-Language-Action Model with Mamba SSM. arXiv:2406.04339. [arXiv]
127. [128] Xu, J. et al. (2025). GeRM: A Generalist Robotic Model via Foundation Models. arXiv:2403.13358. [arXiv]
128. [129] Chen, Q. et al. (2025). PD-VLA: Parallel Decoding for Efficient VLA Inference. arXiv:2503.02310. [arXiv]
129. [130] Wang, X. et al. (2025). Spec-VLA: Speculative Decoding for Accelerating VLA Models. arXiv:2507.22424. [arXiv]
130. [131] Yang, Z. et al. (2025). EgoVLA: Egocentric Vision-Language-Action Model. arXiv:2507.12440. [arXiv]
131. [132] Hung, Y. et al. (2025). NORA: Normalizing Flow-based Robot Action Generation. arXiv:2504.19854. [arXiv]
132. [133] Budzianowski, P. et al. (2025). EdgeVLA: Efficient VLA Deployment on Edge Devices. arXiv:2507.14049. [arXiv]
133. [134] Kim, D. et al. (2025). DiVLA-2B: Diffusion VLA at 2B Scale. arXiv:2412.03293. [arXiv]
134. [135] Park, J. et al. (2026). HyperVLA: Dynamic Policy Generation via Hypernetworks. arXiv:2510.04898. [arXiv]
135. [136] Liu, Q. et al. (2026). AutoQVLA: Automated Quantization for VLA Models. arXiv:2602.03782. [arXiv]
136. [137] Zhang, S. et al. (2025). Humanoid-VLA: Vision-Language-Action for Humanoid Robots. arXiv:2502.14795. [arXiv]
137. [138] Li, W. et al. (2025). Being-H0: Humanoid Robot Foundation Model. arXiv:2507.15597. [arXiv]
138. [139] Chen, X. et al. (2025). FP3: Foundation Policy with Predictive Planning. arXiv:2503.08950. [arXiv]
139. [140] Li, J. et al. (2025). SafeAuto: Safety-Aware Autonomous Driving with VLA. arXiv:2503.00211. [arXiv]
140. [141] Wei, H. et al. (2025). LangCoop V2V: Language-based Cooperative Driving. arXiv:2504.13406. [arXiv]
141. [142] Wang, D. et al. (2025). CognitiveDrone: VLA for Cognitive Drone Control. arXiv:2503.01378. [arXiv]
142. [143] Zhao, R. et al. (2025). RaceVLA: Vision-Language-Action for Autonomous Racing. arXiv:2503.02572. [arXiv]
143. [144] Cheng, H. et al. (2025). NaVILA: Navigation with VLA. arXiv:2412.04453. [arXiv]
144. [145] Zhang, J. et al. (2024). Uni-NaVid: Unified Navigation with Video Diffusion. arXiv:2412.06224. [arXiv]
145. [146] Liu, M. et al. (2025). Mobility VLA: VLA for Mobile Robot Navigation. arXiv:2407.07775. [arXiv]
146. [147] Li, Z. et al. (2024). RoboNurse-VLA: Robotic Nursing Assistant with VLA. arXiv:2409.19590. [arXiv]
147. [148] Chen, J. et al. (2025). ObjectVLA: Object-Centric VLA Model. arXiv:2502.19250. [arXiv]
148. [149] Lin, K. et al. (2024). ShowUI: Vision-Language-Action Models for GUI Automation. arXiv:2411.17465. [arXiv]
149. [150] Li, H. et al. (2025). RoboArena: A Benchmark Arena for VLA Evaluation. arXiv:2506.18123. [arXiv]
150. [151] Nasiriany, S. et al. (2025). RoboCasa365: Large-Scale Robot Simulation Benchmark. arXiv:2603.04356. [arXiv]
151. [152] Zhang, W. et al. (2025). WorldGym: World Model Training Environments. arXiv:2506.00613. [arXiv]
152. [153] Huang, Z. et al. (2025). TactileVLA: Tactile-Enhanced VLA for Dexterous Manipulation. arXiv:2507.09160. [arXiv]
153. [154] Wang, Y. et al. (2025). OmniVTLA: Omni Vision-Tactile-Language-Action Model. arXiv:2508.08706. [arXiv]
154. [155] Tang, Y. et al. (2023). SayTap: Language to Quadrupedal Locomotion. arXiv:2306.07580. [arXiv]
155. [156] Wang, L. et al. (2024). HPT: Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers. arXiv:2409.20537. [arXiv]
156. [157] Physical Intelligence (2025). π^*_0.6: a VLA That Learns From Experience. arXiv:2511.14759. [arXiv]
157. [158] Physical Intelligence (2026). MEM: Multi-Scale Embodied Memory for Vision Language Action Models. arXiv:2603.03596. [arXiv]