APRL Research Philosophy

Purpose and Structure of This Document

This is not a promotional summary of APRL's research areas. It is a textbook-style explanation for reading the trajectory of past work and the current research plan as one logical structure. It addresses three central questions.

1. What problem has APRL repeatedly expanded through its past research?
2. What definition of spatial intelligence unifies those problems?
3. Why do memory, human intent, runtime reasoning, resilience, and multi-robot sharing necessarily follow from that definition?

The argument proceeds as follows.

• Chapter 1: APRL's central problem beyond existing slogans and research categories
• Chapter 2: The research lineage from Scan Context to SSI
• Chapter 3: The relationship among maps, experience, memory, and belief
• Chapter 4: The lifecycle of spatial understanding
• Chapter 5: Task-dependent knowledge and decision relevance
• Chapter 6: Verification through action and belief revision
• Chapter 7: Resilience beyond robustness
• Chapter 8: The relationship between learned weights and runtime intelligence
• Chapter 9: How human intent determines the relevance of spatial knowledge
• Chapter 10: Multiple robots and collective spatial experience
• Chapter 11: APRL's research program in comparison with related laboratories
• Chapter 12: APRL's definition of autonomy
• Chapter 13: Four philosophical foundations
• Chapter 14: A hierarchy of definitions for research pages and official descriptions

Chapter 1. APRL's Central Problem Runs Deeper Than Three Adjectives

Among the suggested readings on the opening page, APRL Research Vision summarizes the current research vision as Persistent / Intent-Grounded / Resilient. Persistence is the ability to maintain knowledge across changes in time and environment; intent grounding is the ability to use human purposes as criteria for judgment; and resilience is the ability to continue a mission after errors and failures. These three qualities are important, but by themselves they do not sufficiently distinguish APRL's research program.

Autonomous operation in complex and dynamic environments, decision-making under long-term uncertainty, alignment with humans, and robust perception and navigation are goals shared by many excellent robotics laboratories. Distinctiveness therefore does not come from owning particular adjectives. APRL's distinctiveness lies in the research object and causal relationships under which these goals are organized.

When APRL's work is arranged along a timeline, a consistent problem emerges beneath the individual technologies.[9] It is not enough for a robot to represent the world accurately once. A robot must build understanding from imperfect observations, judge when that understanding remains valid, select the past experiences needed for the present task, act in alignment with human intent, and revise its understanding through the outcomes of action.

APRL's central problem can therefore be defined as follows.

Spatial intelligence is the capacity to transform experience into reliable action and feed the outcomes of action back into better understanding.

In this definition, Persistent, Intent-Grounded, and Resilient are not independent slogans but necessary conditions for completing a single cycle. If experience does not persist, it cannot inform present action. If it is not aligned with human intent, the robot cannot determine which information matters. If understanding cannot be revised after failure, the outcome of action cannot accumulate as the next experience.

1.1 A List of Research Areas vs. a Research Philosophy

The Situated Spatial Intelligence material classifies APRL's research pillars as GSI (Geometric Spatial Intelligence), RGP (Reality-Grounded Perception), ASM (Agentic Spatial Memory), CSI (Communicative Spatial Intelligence), and ESI (Executable Spatial Intelligence). This taxonomy is useful for describing research areas, but it does not automatically explain why all five should coexist within one laboratory. A first-time visitor may read it as a laboratory that pursues geometry, lifelong mapping, HRI, and embodied AI in parallel.

A research philosophy must state the generative principle beneath the list. For APRL, that principle is the following closed loop.

Model Space → Ground Reality → Remember Experience → Align Intent → Execute and Verify → Revise Understanding

Each research pillar performs one function in this loop. The research areas are therefore not an arbitrary collection of technologies, but components logically required to realize a living spatial understanding.

Chapter 2. Research Lineage: From Maps to Experience

APRL's research philosophy is not a recent narrative projected backward onto earlier work. Read chronologically, the major studies show an expanding scope: from recognizing places, to preserving experience, integrating heterogeneous experiences, selecting experiences relevant to human purposes, and revising understanding from the outcomes of action.[2]

2.1 Scan Context: Recognizing a Place Again

The central question of Scan Context is how a robot recognizes a place when it encounters that place again. This problem is broader than geometric registration alone. The robot must identify structures that persist between present and past observations and connect its current experience to an earlier one.

Place recognition is the most fundamental function of experience-based spatial intelligence. If a current observation cannot be related to a past experience, neither long-term memory nor the reuse of experience is possible.

2.2 1-Day Learning and 1-Year Localization: Change Over Time

Asking whether a representation learned in one day can still be used a year later expands place recognition into a problem of long-term operation. To identify a place as the same after seasons, lighting, structures, and dynamic objects have changed, a robot must distinguish superficial differences in observation from the persistent identity of the place.

At this stage, a map can no longer be an exact copy of a single moment. Long-term localization becomes a problem of learning and judging what has changed and what has remained.

2.3 LT-Mapper: Maintaining a World Across Multiple Times

Rather than forcing worlds observed at different times into one fixed map, LT-Mapper points toward representing states and changes over time together. It expands mapping from the production of a single artifact into long-term state management.

For a robot to operate persistently in a changing environment, deleting past states or overwriting them with the present is not enough. The robot must judge when and under what conditions a state was observed, whether a difference is temporary or structural, and which state is relevant to the present task.

2.4 HeLiPR: Connecting Heterogeneous Sensor Experiences

Different LiDARs and platforms experience the same space with different resolutions, fields of view, and noise characteristics. The problem addressed by HeLiPR is not merely sensor conversion, but the construction of common spatial relationships across heterogeneous experiences.

This problem becomes a foundation for later multi-robot experience sharing. Because experience is not pure information independent of sensor and embodiment, it must be connected while accounting for its provenance and observation conditions.

2.5 ScaleMaster and MR.ScaleMaster: Integrating Multiple Visual Experiences into a Metric World

Each monocular visual sequence may have a different scale ambiguity. Integrating multiple experiences into one metric world requires reasoning about not only the internal structure of each experience, but also the scale and relationships among experiences.

This work expands foundation geometry from reconstructing a single image to consistently integrating multiple experiences. What matters is not only extracting as much geometry as possible from one observation, but combining partial understandings created from different observations into one actionable world model.

2.6 LT-Mem: Which Past Should Be Used Now?

As memory grows, retrieval becomes harder. The most similar past experience is not necessarily the most useful one for the present decision. The core question raised by LT-Mem is which memory should be trusted and used in the current situation.

Here, long-term memory is not merely a repository. It is a component of decision-making in which the recency and conditions of an experience, its success or failure, its agreement with current observation, and its expected action value must all be considered.

2.7 Implicit Goals and VLN: Selecting Experience Relevant to Human Intent

Recent work on implicit-goal navigation and vision-language navigation introduces human intent into the retrieval conditions for spatial memory. Human instructions may not arrive as complete coordinates or explicit goals. A robot must interpret the situation, preferences, and constraints embedded in language, and then find the spatial experiences relevant to that intent.

At this stage, APRL's work expands from recognizing places and maintaining maps to deciding which spatial knowledge should be used to produce the physical outcome a human wants.

2.8 Convergence Toward Experience 2.0 and SSI

The 2024 research plan reinterpreted SLAM not merely as a tool for estimating pose and map, but as an Automated Experience Reconstruction Machine. It also proposed Experience 2.0 as a research object that combines experiences generated across multiple robots, sensors, and times, beyond a single journey.

Subsequent work on spatial memory, implicit human intent, runtime verification and recovery, and Situated Spatial Intelligence further generalizes this direction. The lineage can be summarized as follows.

Place Recognition → Lifelong Localization → Multi-Temporal Mapping → Heterogeneous Experience → Metric Integration → Spatial Memory → Human Intent → Runtime Verification and Recovery → Situated Spatial Intelligence

The object running through this lineage is not the map, but the formation, connection, selection, execution, and revision of experience.

Chapter 3. Maps, Experience, Memory, and Belief

Understanding APRL's spatial intelligence requires distinguishing maps, experience, memory, and belief. The four concepts are connected, but they are not the same.

3.1 A Map Is a Compression of Spatial Information

A map primarily represents what is located where. It structures locations, shapes, connectivity, semantic categories, traversability, and related information for use in planning and localization. A map is essential to a robot's spatial judgment, but it does not contain the entire real world as it is.

A map contains unobserved regions, sensor errors, temporal changes, and omissions introduced by representation. The information needed from the same space also differs by task. A map is therefore not the world itself, but a model constructed under particular purposes and observation conditions.

3.2 Experience Includes Conditions, Actions, and Outcomes

Experience is broader than a map. It includes not only what was seen, but when it was observed, through which sensor and embodiment, for what purpose, what action was taken, whether it succeeded or failed, and what costs and risks occurred.[4]

For example, the statement “there is a door at this location” may belong in a map. But “it was open on weekday mornings, a robot of a particular size could pass through it, and the latest attempt found it locked and required a detour” is closer to the structure of experience.

3.3 Memory Makes Experience Usable in Present Decisions

Memory is not simply the storage of past experience. It is a system organized to retrieve experiences relevant to the current situation, assess their reliability, and use them in action decisions. The quality of memory cannot be evaluated only by the number of stored items or by recall accuracy.

The ultimate evaluation question is: Did using that memory improve the robot's present action? Memory is therefore not an auxiliary database for perception, but a decision mechanism that improves action quality.

3.4 Belief Is the Provisional Understanding Available Now

A robot combines current observations, past experience, and learned priors to form beliefs about the world. A belief is not a confirmed fact, but the state judged most plausible under the available evidence. It must be revised when new observations or action outcomes arrive.

This is where the following sentence becomes meaningful.

The map is a hypothesis, not the truth.

This is not APRL's final definition, however. It is an epistemological motivation for why cyclical spatial intelligence is necessary. Because the map is a hypothesis, a robot must check its confidence, use memory conditionally, test it through action, and revise it according to outcomes.

3.5 What It Means to Know a Place

Knowing a place cannot mean only reconstructing its geometry. To claim that a robot knows a place in a long-term, real-world setting, the robot must at least be able to answer the following questions.

• When and how does the place change over time?
• Under what conditions have particular actions succeeded or failed?
• Which past experiences remain valid now?
• Which information has become stale and should be observed again?
• Which memories are needed for the present task?
• How confident is the robot in its judgment?
• Under uncertainty, what should it observe again, or whom should it ask—and what?

This definition goes beyond lifelong mapping. The goal is not merely to create a map that lasts, but to use long-accumulated experience appropriately in present action.

Chapter 4. The Lifecycle of Spatial Understanding

APRL's research program can be organized as a lifecycle in which spatial understanding is created, used, and revised.[3]

4.1 Construct: Build Understanding from Imperfect Observations

A robot infers geometry, objects, places, relations, and affordances from sensor observations. Because observations are partial and depend on sensors and viewpoints, the resulting understanding contains uncertainty from the outset.

4.2 Ground and Verify: Ground the Model in Reality

The robot must confirm that the constructed model agrees with present reality. This process includes localization, data association, loop closure, change detection, cross-sensor alignment, and confidence estimation.

4.3 Remember and Update: Maintain Experience Over Time

A robot cannot store every observation in the same way. It must decide what to preserve, which experiences to merge, how to retain conflicting experiences, and when to update stale information.

4.4 Query and Select: Choose the Experience Needed for the Present Decision

Retrieval cannot be based on similarity alone. Experience must be selected according to the present task, risk, information recency, expected action value, and human intent.

4.5 Align and Communicate: Connect Human Intent to Spatial Knowledge

Human language does not arrive as complete coordinates. A robot must interpret unexpressed preferences and constraints, and clarify intent through questions when uncertainty is high.

4.6 Execute: Transform Understanding into Physical Outcomes

Spatial understanding becomes real action through planning, navigation, manipulation, and interaction. Sensors, computing, embodiment, dynamics, and safety constraints converge at this stage.

4.7 Verify and Revise: Update Understanding from Action Outcomes

The success, failure, cost, and gap between expected and actual outcomes are new evidence. A robot uses this evidence to revise its maps, memories, beliefs, and planning strategies.

The overall structure is as follows.

World → Observation → Experience → Memory → Belief → Intent → Action → Outcome → Belief and Memory Revision

In this structure, spatial intelligence is not a representation but a continuously operating epistemic control loop.

Chapter 5. Knowledge Is Task-Dependent

One of APRL's central propositions is that a good world model is not necessarily the model that represents the world in the greatest detail. A good world model preserves information needed for the current purpose, retrieves it at the appropriate moment, assesses its reliability, and makes it available for action.

5.1 The Same Space, Different Knowledge

The information required from the same building differs across home assistance, hospital guidance, delivery, and rescue missions. A home-assistance robot may prioritize user preferences and daily patterns; a hospital guide robot, accessibility, congestion, and operating hours; a delivery robot, traversability and time cost; and a rescue robot, hazardous zones and survival likelihood.

Storing and processing every fact about the world at the same level is not only computationally inefficient, but also unhelpful for judgment. The task determines what should be observed, remembered, retrieved, and verified.

5.2 Semantic Similarity and Decision Relevance

Linguistic or visual similarity is useful when retrieving past experience, but it is not sufficient. The present judgment may require a record relevant to the action decision rather than the most semantically similar record.[4]

Decision relevance may include the following factors.

• Is the current route actually traversable?
• How old is the information?
• Does it agree with current sensor observations?
• Would using this experience reduce risk or cost?
• Does it apply to the current robot's embodiment and sensor conditions?
• Does it satisfy the present human intent and constraints?

APRL's question therefore shifts from “What is the true representation of the world?” to “What must this agent know in order to act correctly now?” This is a pragmatic definition of spatial intelligence.

5.3 The Ultimate Metric for Memory

High retrieval accuracy alone is not enough when evaluating a memory system. If retrieved memories induce a bad plan or recommend a stale route, high recall does not lead to good action.

The ultimate performance of memory should be connected to action-level metrics such as:

• success rate and task-completion time;
• risk and energy cost;
• reduction in unnecessary re-exploration;
• frequency of human intervention and clarification;
• recovery time after failure; and
• utility of experience transferred to another robot.

From this perspective, memory is not the ability to reproduce the past well, but the ability to improve present and future action.

Chapter 6. Action Is a Means of Verifying Understanding

Traditional pipelines place perception, reasoning, and action in sequence. In this structure, action appears to be the result of already completed perception and reasoning.

Perception → Reasoning → Action

In the real world, however, action is both the use of understanding and a test of that understanding.

Perception → Belief → Action → Outcome → Belief Revision

6.1 Why Action Outcomes Become Evidence

Suppose a robot concludes from memory that a door is open and chooses a route through it. If the door will not open on site, this is not merely a failure of the navigation module. It is evidence that a particular belief in the robot's world model is no longer valid.

From this outcome, the robot must distinguish among multiple hypotheses: whether the door is temporarily locked, operating hours have changed, localization is wrong, another door was recognized as the same place, or passage is impossible only for its particular body.

6.2 The Robot as a Scientific Agent

In this structure, robot operation resembles the following cycle.

Hypothesize → Act → Test → Revise

A robot forms a hypothesis from current evidence, tests it through action, and uses the outcome to revise its understanding. Action does not exist only downstream of perception; it is an active experiment that produces new perception.

6.3 The Meaning of Failure Records

Failure is not a log to discard, but an important part of experience. Preserving the conditions under which a plan failed can prevent the same cost from being paid again in the next action. When the conditions of both success and failure are remembered, spatial memory develops from scene memory into knowledge of actionability.

Chapter 7. Resilience Is Broader Than Robustness

Robustness is the ability to reduce performance degradation and the likelihood of failure under given disturbances and uncertainty. Resilience is the ability to diagnose a state, revise a strategy, and continue the mission after failure has already occurred or an existing assumption has collapsed.[5]

7.1 Stages of a Resilient System

A resilient robot must carry out the following functions cyclically.

1. Detect: Detect its own uncertainty, errors, and failures.
2. Diagnose: Estimate the cause and scope of a failure.
3. Acquire: Observe again or ask a human for the required information.
4. Replan: Revise beliefs, memories, plans, and action strategies.
5. Resume: Continue the mission with the revised strategy.
6. Learn: Preserve the outcome as experience for the next decision.

This cycle is not mere exception handling. It is an epistemic process that uses failure as an opportunity to update knowledge.

7.2 Systems That Never Fail and Systems That Handle Failure

No learned model can include in advance every sensor failure, environmental change, ambiguous human instruction, or exceptional physical situation that may occur in an open world. The core of real autonomy is therefore not the complete elimination of failure, but the ability to detect, understand, and recover from it.

Resilience matters at APRL because it connects persistence and intent grounding to real operation. Unconditionally trusting old memories makes persistence dangerous, while failing to clarify ambiguous intent can produce the wrong action. Resilience is the execution structure that makes persistence and intent alignment safe to operate.

Chapter 8. Intelligence Does Not Exist Only in Learned Weights

Modern embodied-AI and VLA research internalizes more behavioral capability in models through large-scale data, foundation models, imitation learning, and reinforcement learning. This is powerful, but it does not mean that every situation and failure a physical robot may encounter can be compiled into model weights in advance.

8.1 Priors and Runtime Evidence

A learned model provides priors about the world and action. In real deployment, however, current sensor observations, past experiences, human instructions, system states, and action outcomes continue to arrive. A robot must combine priors with runtime evidence to form its current belief.

The 2024 research plan viewed robot intelligence as a Bayesian update of priors with runtime sensor evidence. That perspective can now be extended to memory, reasoning, verification, interaction, and recovery.

8.2 The Meaning of Runtime Harnessing

Planners, memory, verifiers, and uncertainty detectors should not necessarily be treated as temporary aids that will disappear when a larger end-to-end model arrives. In an open world, the structure that coordinates these functions at runtime may itself be part of intelligence.[6]

This can be expressed conceptually as follows.

Intelligence = Learned Prior + Current Observation + Experience + Runtime Deliberation + Verification

This is not a mathematical model that simply adds the terms. It is a structural proposition that real intelligence exists in the combination of a learned model and the process of execution.

8.3 Model Capability and System Capability

A model provides capabilities for perception, prediction, language understanding, and action generation. A system decides when to trust the model, which memories to retrieve, when to plan, what to verify, when to ask a human, and how to recover after failure.

APRL therefore addresses both the problem of building better models and the problem of organizing their capabilities into reliable action in the real world.

Chapter 9. Humans Determine Relevance

At APRL, a human is not merely a command-input device or a user interface. Human intent is a condition that determines which elements of accumulated spatial knowledge matter now.

9.1 From Language Understanding to Spatial Decision-Making

Consider the statement, “My leg hurts a little, and I need to go to another floor.” It does not directly specify a goal coordinate. A robot must nevertheless infer the following.

• Where does the user intend to go?
• Is there an intention to avoid stairs?
• Where is the elevator, and what is its current operating state?
• Has the robot experienced congestion at a particular time in the past?
• What route and speed fit the user's condition?
• What should the robot ask to resolve uncertainty?

This problem cannot be solved by NLP alone. It is a spatial decision problem combining language, spatial memory, current observation, human state, mobility, and uncertainty resolution.

9.2 The Role of Language

At APRL, language can be understood as an interface for querying and constraining spatial intelligence. Human expression determines which memories to retrieve, which routes to avoid, which risks to prioritize, and when to request clarification.

9.3 The Precise Meaning of Intent-Grounded

Intent-Grounded does not simply mean following linguistic commands well. It is the ability to use a human purpose as the criterion for physical outcomes, select the spatial knowledge required for that purpose, and clarify intent through further interaction when uncertainty remains.

Chapter 10. Multiple Robots Create Collective Spatial Experience

A common description of traditional multi-robot SLAM is the registration and merging of maps from several robots. From APRL's experience-centered perspective, the goal is broader.

One robot should not have to pay again the cost of an experience already paid by another.

10.1 Experience Cannot Be Copied as Is

Robots differ in sensors, fields of view, height, mobility, payload, and safety constraints. A route traversable by one robot may be impossible for another, and a landmark reliable to one sensor may be invisible to another.

The following information must therefore accompany shared experience.

• Provenance: Who had the experience, when, and under what conditions?
• Validity: Does the experience remain valid now?
• Embodiment: With what body and mobility was the outcome obtained?
• Sensor context: With which sensors and observation range was it constructed?
• Outcome: Which action succeeded or failed, and at what cost?

10.2 A Collective Prior

Experience whose provenance and conditions are preserved can become a prior for another robot. A new robot can choose observations and actions by drawing on other robots' successes and failures instead of exploring every space from scratch.

This concept can be called collective spatial experience. The 2024 ideas of the Robot Web and Experience 2.0 connect to a direction in which the experiences of multiple agents accumulate as a collective prior.

Chapter 11. APRL's Distinctive Position

APRL is not the only laboratory to study the individual subproblems it addresses. Distinctiveness lies not in monopoly over a problem, but in the intersection and organization of the research program.

APRL's distinctive position lies in carrying the research question all the way from geometry to memory, from memory to human intent, and from human intent to runtime reasoning and physical outcomes.

In other words, APRL does not study only “how to represent space accurately.” It treats “how experience of space is used in action, and how the outcome of action becomes knowledge again” as one continuous research object.

Chapter 12. Autonomy Is Not the Absence of Humans

For APRL, autonomy does not mean removing humans from the system. It is the ability to retain human intent as the goal without requiring a human to perform perception, planning, error diagnosis, and recovery at every moment.

Autonomy is not the absence of humans.

It is the ability to fulfill human intent without requiring continuous human intervention.

This definition clarifies the relationship between intent grounding and minimal human intervention.

• Intent grounding defines what must be achieved.
• Minimal human intervention defines how independently and persistently that goal is achieved.
• Resilience defines how the goal is maintained through failure and uncertainty.

The ultimate metric for autonomy is therefore not only a score on a particular benchmark. An important criterion is how long a robot can produce safe and accurate physical outcomes without continuous human intervention.[5]

Chapter 13. Four Philosophical Foundations

APRL's research program can be interpreted as the combination of four philosophical perspectives within spatial intelligence. APRL did not invent the perspectives themselves, but its position is formed by the way it connects geometry, memory, human intent, runtime reasoning, and physical outcomes in one closed loop.

13.1 Bayesian Epistemology

A robot's knowledge of the world is not absolute truth, but belief continually revised by evidence. Current observations, past experiences, learned priors, and action outcomes all contribute to belief updates.

This perspective does not end with acknowledging uncertainty in maps and memory. It extends to the entire process of representing uncertainty, choosing observations and actions that obtain new evidence, and updating beliefs.

13.2 Pragmatism

The final criterion for a good representation is not reconstruction score alone. What matters is whether the information improves actual action, reduces risk and cost, and better fulfills human purposes.

13.3 Situated and Enactive Cognition

Intelligence is not a static property contained inside a model. It is enacted and revised through body-world-human interaction. A robot's body and environment are not external conditions of intelligence, but parts of the process through which intelligence exists.

13.4 Distributed Cognition

Intelligence does not exist only in the internal state of one robot. A collective prior forms when the experiences of multiple robots and humans are shared while preserving provenance and embodiment.

13.5 Combining the Four Perspectives

The four perspectives connect as follows.

1. Bayesian epistemology treats knowledge as revisable belief.
2. Pragmatism evaluates the value of belief by action outcomes.
3. Situated cognition treats action and interaction as part of belief formation.
4. Distributed cognition transfers formed experience among multiple agents.

This combination distinguishes APRL's “spatial intelligence” from spatial representation alone.

Chapter 14. Structuring the Research Program and Its Official Description

Research pages and research statements should distinguish sentences that operate at different levels. Combining definition, motivation, method, and value judgment in one sentence weakens the message.

14.1 Level 1: The Laboratory's Central Definition

The opening sentence should positively define what APRL studies.

APRL studies how experience becomes reliable action—and how the outcomes of action return as better understanding.

This sentence connects experience, reliable action, and feedback to understanding in a bidirectional structure. Mapping, memory, intent, runtime verification, and recovery can all follow from it.

14.2 Level 2: An Academic Definition of the Research Object

APRL studies the lifecycle of spatial understanding.

This sentence names the research object as the lifecycle of spatial understanding. It is the most stable academic definition for use in papers, research plans, and talks.

14.3 Level 3: Operating Principle

A robot's understanding of the world is never finished. It is constructed from imperfect observations, maintained across change, aligned with human intent, tested through physical action, and revised from experience.

This paragraph explains the stages that make up the lifecycle.

14.4 Level 4: Epistemological Motivation

The map is a hypothesis, not the truth.

This belongs not as the opening sentence, but as the motivation for continuous verification and revision.

14.5 Level 5: Research Pillars

Placing the following cycle beneath the central definition makes GSI, RGP, ASM, CSI, and ESI read not as a parallel taxonomy, but as a program derived from a single research philosophy.

Model Space → Ground Reality → Remember Experience → Align Intent → Execute and Verify → Revise Understanding

14.6 Supporting Lines

The following lines can serve supporting roles according to context.

We study how spatial understanding lives.

Spatial intelligence is a living understanding shaped by experience and revised through action.

From Maps to Experience. From Experience to Action. From Action to Understanding.

These lines are better used as presentation titles, chapter headings, or diagram captions than as replacements for the central definition.

Conclusion. APRL Studies a Living Spatial Understanding

The most coherent way to describe APRL's research is not to list individual technologies or three adjectives. Its past research began with recognizing places and expanded to change over time, heterogeneous sensor experience, metric integration, long-term memory, human intent, runtime reasoning, recovery after failure, and multi-robot experience sharing.

At the center of this trajectory is one cycle. A robot builds spatial understanding from observations, selects past experience for its present purpose, aligns it with human intent, tests its beliefs through action, and stores the outcome as new experience.

APRL's spatial intelligence is therefore not equivalent to a well-drawn map of the world. It is a living, self-revising process that transforms experience into action in a changing world and returns action to understanding.

APRL studies how experience becomes reliable action—and how the outcomes of action return as better understanding.

This definition connects the research trajectory from Scan Context through lifelong localization, LT-Mapper, heterogeneous mapping, Experience 2.0, ScaleMaster, LT-Mem, implicit human intent, runtime verification and recovery, and SSI within one logic. It also provides a criterion for evaluating future work: whether a new study belongs to APRL's program can be judged by the role it plays in better constructing, remembering, selecting, aligning, testing, and feeding back spatial understanding into the next experience.

References

1. Autonomous Systems Lab, ETH Zurich
2. APRL Publications
3. APRL's Research Program and Situated Spatial Intelligence
4. Should Robots Remember Maps, or Should They Remember Experience?
5. Redefining the Work of Robotics Research
6. Should Robots Learn More, or Think More?
7. Stanford ILIAD - Human-AI/Robot Interaction
8. MIT SPARK Lab
9. APRL Research