Form follows function.
Atlas는 왜 이렇게 생겼나

영상에 포함된 모든 대화 내용을 빠짐없이 기록한 자막입니다.

0:00 Thanks for joining us. I'm Chris Thorne.
0:02 I'm here with Aaron Abbruff and James
0:04 Cuso. We'll be discussing the new Atlas
0:06 hardware design today. I lead the Atlas
0:09 hardware team. I've been at Boston
0:10 Dynamics for about 15 years. Over half
0:13 of that was working on Atlas
0:15 exclusively. Uh prior to joining BD, I
0:19 was at the Grasp Lab at the University
0:21 of Pennsylvania where I received my PhD
0:23 in mechanical engineering. Aaron, can
0:25 you give us a little bit of background
0:26 on yourself? I've been working with
0:28 Boston Dynamics for six years. Uh I
0:31 originally came on to lead the
0:34 industrial design for the stretch
0:35 program. Prior to that, I worked for
0:38 design agencies for about 25 years. Um
0:42 working in medical devices for the first
0:45 part of that and then eventually
0:48 transitioning over to consumer products.
0:50 At the tail end of that work, I led the
0:53 design on the spa program.
0:55 Great. James, can you tell us a little
0:57 bit about yourself and how you came to
0:58 Boston Dynamics?
0:59 Sure. I just joined BD about nine months
1:03 ago. Um, prior to that, I spent first
1:06 part of my career in automotive and
1:09 right before this, I was in additive
1:11 manufacturing,
1:12 but in between that I spent uh 12 years
1:16 in consumer electronics
1:18 um at Apple where I led product design
1:21 for the Mac division. So, a lot of
1:24 people that work at Boston Dynamics love
1:25 to talk about their first time seeing
1:27 the robots in real life. So, I thought
1:29 it'd be fun to see kind of what your
1:31 guys experience were. I'm going to talk
1:33 about the first time I saw Atlas cuz
1:35 that was truly memorable. You know,
1:37 there's lots of videos of Atlas on
1:39 YouTube and whatnot, and I've seen them
1:41 all, but when you're watching a video on
1:43 YouTube, something in the back of your
1:45 mind is telling you like, "This isn't
1:46 real." You know, this is AI. When you're
1:48 in the lab and you see Atlas standing in
1:51 front of you doing its thing, your brain
1:54 at first wants to say like, "This can't
1:56 be real." But then you realize, "No,
1:58 this is right in front of me." Like,
1:59 "This is as real as it gets." And like,
2:01 I'm just getting goosebumps now thinking
2:03 of that moment because it was, you know,
2:05 I don't have that anymore.
2:08 And that's kind of a shame because it's
2:10 just so natural to me to be around these
2:12 robots. But I'll never forget that first
2:14 day.
2:15 How about you? The first time I saw a
2:18 Boston Dynamics robot in person was uh
2:21 the hydraulic spot. And at this point, I
2:24 have a hard time remembering, you know,
2:26 seeing it in person versus seeing it on
2:29 YouTube, having it, you know, be kicked
2:32 by somebody at Boston Dynamics. But the
2:34 first time I ever saw a Boston Dynamics
2:37 robot and was made aware of the company
2:39 was in the early 90s where I saw one of
2:42 the two-legged robots uh on a video and
2:47 it was uh the most incredible thing I'd
2:50 ever seen. You know, it was it seemed
2:52 like uh the world was going to change
2:54 and you know this was right around the
2:56 corner, you know, and uh you know, I
2:59 never expected that I would be working
3:01 for the company one day. Yeah, I think
3:03 for me the first time I saw the robots
3:05 or some of them anyway was uh during my
3:08 interview and you know I did my
3:10 presentation they took me through the
3:12 lab to look at I think I saw Sanfle LS3
3:16 um cheetah and maybe a few others and
3:21 just thinking it was like the coolest
3:22 thing I've ever seen and we when we were
3:25 doing it they were asking me a ton of
3:26 questions about the robots like why do
3:28 you think we did this or you know how do
3:29 you think this works and I didn't even
3:31 realize at at the time that it was still
3:33 part of the interview. And then I went
3:34 back to my hotel room and talked to my
3:36 wife and I said, "I can't believe this
3:39 is a real job. Like, I can't believe
3:40 people get paid to do this. Like, I need
3:42 to work there. This is literally a dream
3:44 job for a mechanical engineer." So,
3:46 Aaron, when you first arrived at BD, we
3:49 had just started our spot product
3:50 journey and you were just starting to
3:52 think about what our design language was
3:54 for our products. How has your thinking
3:55 in that area evolved over time and how
3:58 does it apply to the new Atlas design?
4:01 So when I first came on to BD, I'd been
4:04 working with Spot for three years. When
4:07 we first started that program, the real
4:09 challenge was how do we make this
4:12 otherworldly robot look like a product?
4:16 That's really what our design goal was.
4:19 And uh you know the challenge there is
4:21 uh a lot of companies um a lot of people
4:24 really want to make these robots look
4:26 like the robots that they grew up with.
4:28 you know, what they what they expect a
4:31 robot to look like, which usually is
4:33 something from a science fiction, you
4:35 know, a movie prop or a costume. And
4:38 that's not really delivering on what a
4:40 product needs to be. So, when we were
4:43 thinking about spot, we were thinking
4:46 about, well, how's this robot going to
4:47 be used? And we didn't quite know, and
4:51 we were still exploring all of the
4:52 possibilities. So, we're thinking about
4:55 this as a, you know, really a modular
4:58 platform that our customers could use,
5:02 you know, as they saw fit and and
5:03 allowed them to put their products on on
5:06 the back of it, whether that's
5:07 instrumentation or something that would,
5:10 you know, measure something during
5:12 inspection. We were really looking at
5:15 how the robot moved and trying to
5:18 reflect that movement in the design of
5:20 the robot. So, a lot of the cladding
5:23 that you see on spot is there to protect
5:26 the robot while it's navigating its
5:28 environment. Whether it's going upstairs
5:30 or, you know, falling downstairs, it
5:33 needs to be able to protect its cameras
5:35 and the sensitive instrumentation. The
5:38 stretch robot is a much more traditional
5:41 robot and it's a purpose-built robot
5:44 going into a customer environment in
5:47 this case uh a shipping center or
5:50 warehouse. Stretch needs to look like
5:53 similar equipment like fork trucks and
5:55 equipment that people are used to
5:57 working around. Atlas is a purpose-built
5:59 robot that we're putting into an
6:01 environment with people and that's a
6:03 manufacturing environment and we're
6:06 prioritizing the tasks that the robot is
6:08 going to be doing. We're targeting a
6:10 humanoid capability, but we're not
6:12 targeting a humanoid form.
6:14 So, James, uh, you started recently at
6:17 the company and, uh, what did you think
6:19 the first time you saw the concept
6:20 sketches for the new robot?
6:23 Yeah. So, I remember um our first
6:25 meeting after I was hired, but before I
6:27 started, we met in your office and on
6:30 your wall is a life-size print out of
6:35 the Atlas robot. And then you started to
6:38 talk through some of like the guiding
6:41 principles behind the robot, you know,
6:44 the 247
6:45 uptime, the desire to have continuous
6:49 range of motion on the joints, the
6:51 desire for it to be robust, the desire
6:55 to improve serviceability by having a
6:57 lot of reused parts that could quickly
7:01 be swapped on and off in the event
7:03 something got damaged. And as you
7:05 explained that, the robot just makes
7:07 sense.
7:09 You know, one example is its ability to
7:11 change its own batteries, right?
7:13 Right.
7:13 Rather than hide the batteries beneath
7:15 the skin, the batteries are prominently
7:18 placed on the robot because the robot
7:20 needs to be able to access them. And so
7:22 we leaned into that.
7:24 Yeah, the runtime was really a big
7:26 product requirement, right? especially
7:27 for the industrial use case because we,
7:29 you know, we did a lot of thinking about
7:32 could we get away with an internal
7:34 battery, you know, and and if we
7:36 leverage fast charging technology, you
7:38 know, could you have a fleet of robots
7:40 in a factory that could go fast charge?
7:42 And then you start to work through it
7:44 and realize that that would be a lot of
7:46 power consumption to have a fleet of
7:48 robots that need to fast charge. And the
7:51 continuous runtime was such a priority
7:53 especially for Hyundai, you know,
7:55 because eventually they want to go into
7:57 general assembly and there is really no
7:58 downtime there. They need robots working
8:00 24/7. And so it became pretty clear that
8:04 swapping its own battery was where we
8:06 had to go. And then of course having two
8:09 batteries means you can always have the
8:10 robot operational while it's swapping
8:13 its battery. So that solution just kind
8:15 of presented itself as like where we had
8:17 to go to to really make a compelling
8:19 industrial robot.
8:20 When I look at the robot, you know, it's
8:22 obvious to me that it's not a form
8:25 factor that's possible with just, you
8:30 know, your average
8:32 available technologies. And you
8:34 explained to me that like these
8:36 actuators sort of unlock this
8:38 morphology. Can you say more about that?
8:42 Yeah, I think it's definitely the
8:45 actuators that make this robot possible.
8:48 Early on, we invested heavily in
8:51 actuation technology, which I think paid
8:53 off. Uh because the actuators are,
8:56 depending on what metric you're
8:58 interested in, something like two to
9:00 five times more performant than anything
9:03 we could buy uh today. So, that enabled
9:06 us to put the same actuator in a lot
9:09 more places on the robot. When you take
9:11 some of our older robots, we've got
9:13 unique actuators in different locations
9:16 because they package better. Their their
9:18 performance characteristics make sense
9:20 for those locations. But when you have a
9:23 really compact, you know, power dense
9:25 actuator, you can now put the same one
9:27 that you have in the hip as you can in
9:29 the ankle. And you unlock all this
9:31 modularity and simplicity in the robot
9:34 that you just couldn't get any other
9:36 way. So by investing in the actuator, we
9:39 drastically simplified the robot. Most
9:41 of the structures are just simple
9:42 structural pieces connecting actuators
9:44 together.
9:45 From an engineering standpoint, that
9:46 reuse is really great because, you know,
9:50 the hardest part about engineering a
9:52 device is making sure it's really robust
9:55 and reliable. And the only way to do
9:57 that is build a lot of them and find
10:00 every last little problem. And when
10:02 you're building a robot that has 10 of
10:04 one actuator and what 13 of another,
10:07 you're already building a lot of motors
10:09 in just a single robot. You multiply
10:11 that by a few robots and a few more and
10:13 you all a lot of motors. So you're
10:15 learning a lot,
10:17 right? So that reuse not only benefits
10:19 the robot's design, but it benefits the
10:23 engineering learnings, it benefits the
10:24 manufacturing, taking advantage of the
10:26 economy of scale, the sourcing, um, and
10:31 the service strategy, right? Being able
10:34 to stock a single arm that can be
10:36 populated on either side of the robot
10:39 and then being able to stock two types
10:41 of motors that can be, you know, used in
10:44 any one of the joints to take a take a
10:47 limb that has been taken out of service
10:48 and bring that back online. It's such a
10:51 cool approach.
10:53 Well, from a design standpoint, the
10:56 modularity dictates the visual design of
10:58 the robot.
10:59 Mh. you know, you can't get around the
11:02 shoulder is going to look exactly like
11:04 the hip. The upper leg is going to look
11:06 a lot like the upper arm. You know, the
11:09 right leg and the left leg are the same
11:11 part. And the same with the arms. So,
11:13 there's no real front and back to those
11:16 limbs. They're symmetrical, you know.
11:18 And I think one of the biggest
11:20 challenges was if we make a change to
11:23 one of those actuators, it's going to
11:25 impact across the entire robot, you
11:29 know. So if we change uh a dimension, if
11:34 we make it a little bigger, it's just
11:36 going to times four, you know, increase
11:39 the height of the robot. That was the
11:41 biggest challenge probably with uh
11:44 regard to just the commonality across
11:46 the robot.
11:47 The challenge with executing these very
11:50 power dense motors is dealing with the
11:53 thermals, right? They're generating a
11:55 ton of heat.
11:56 Yeah. And that's both an engineering
11:57 problem and a design problem, right? So,
12:00 how did you all approach that?
12:02 Yeah, that one was was tricky. It might
12:04 be one of the things I'm most proud of
12:05 that we were able to accomplish in the
12:07 actuator design. We spent a lot of time
12:09 trying to make it as efficient as
12:10 possible so that we weren't having to
12:13 manage a ton of heat. But like you said,
12:15 you know, either way, you're managing a
12:17 ton of heat. So, we made the decision to
12:19 try to go after passive cooling, which
12:21 would again drastically simplify the
12:23 robot. We don't have to have fans
12:24 everywhere. So there's only one fan in
12:26 the robot and it's in the head. Uh there
12:28 are no fans on any of the actuators. And
12:31 to accomplish that, we had to do a ton
12:33 of analysis work uh to make sure that we
12:36 could passively cool every actuator
12:38 through all the behaviors in all the
12:40 ambient temperature conditions um in
12:43 these industrial settings. And that was
12:44 a huge challenge. But you should speak
12:46 to this, but I think it probably
12:47 contributed more than we originally
12:49 thought to the the visual design of the
12:51 robot.
12:52 Sure. Well, when we started that wasn't
12:55 something that we were focusing on and
12:57 it emerged as a a goal somewhere mid-
13:00 project and I think we had some late
13:03 night discussions uh you know about well
13:06 can we increase the fins uh by you know
13:10 a couple of millimeters and I would I
13:13 would think about well how much taller
13:14 is that going to make the robot and now
13:17 how do we how do we think about pinch
13:20 and and safety because all of these
13:23 things are squeezing together a little
13:24 bit more. We took the cooling fins and
13:28 made it a cosmetic part of the robot,
13:30 right? So when you when you see that on
13:32 the outside of the robot, on the outside
13:34 of the legs, on the outside of the arms,
13:36 that is a functional part of the robot.
13:38 And you know, we're we're encouraging
13:41 air flow. Air flows behind the padding.
13:45 And
13:47 this it's great not having to to deal
13:49 with fans and the noise of fans.
13:52 Yeah. So you mentioned pinch which was
13:53 another really big uh requirement
13:57 specifically around safety handling
13:59 safety you know people handling the
14:01 robot uh in its offstate robot
14:04 interacting with the environment in in
14:06 different ways. How difficult was it to
14:09 incorporate you know pinch safety and
14:11 handling safety into the robot? It adds
14:15 a level of complexity
14:17 that
14:19 just increases the amount of thinking
14:21 that you have to put into every part of
14:23 the robot. Uh and and what we really
14:26 wanted was at least a one inch gap in in
14:30 all these places where we were concerned
14:32 about someone being pinched or uh you
14:35 know worrying about an entrapment. You
14:37 know the challenge also is we're putting
14:40 cooling fins on those surfaces. So, we
14:43 want to create as much clearance as
14:44 possible. It's a hot surface. There are
14:48 fins that we don't want to press
14:50 somebody's hand against. Yeah. The
14:52 safety and the pinch across the entire
14:54 robot uh are something that we've put a
14:57 lot of time into. And, [clears throat]
14:58 you know, it impacts the robot in ways
15:01 that we're not expecting. If you need to
15:03 accommodate a 1-in gap, that's going to
15:07 uh immediately impact the height of the
15:09 robot. you know, we're we're concerned
15:11 about the the pinch between the head and
15:14 the shoulders or the the pinch between
15:17 the pelvis and the midback, you know,
15:20 and it'll impact the width, you know.
15:23 So, we're we're concerned about the
15:24 pinch in the knee and the pinch in the
15:28 elbow. And, you know, to to address
15:32 this, we created offset links.
15:35 Yeah.
15:35 You know, those offset links make the
15:37 robot wider. I think the the legs were
15:40 the biggest concern because they're
15:42 they're an obvious departure from a
15:45 human form.
15:47 And you know, that's that's what
15:50 everybody is expecting to see. And you
15:53 know, we have something very similar on
15:55 the elbows. You know, there's there's an
15:57 offset lower arm to the upper arm. And
16:01 again, you know, this is for safety and
16:03 it's to increase the ROM as much as
16:06 possible. Yeah, I think in general I was
16:08 surprised when we unveiled the robot
16:11 that a lot of people said what you said,
16:13 which is, "Oh, this makes sense." It's
16:15 like something people hadn't seen, but
16:17 they they thought it I I was convinced
16:19 we were going to get a ton of people
16:21 being like, "This robot looks really
16:23 weird. I don't like it. Why doesn't it
16:25 look like all the other ones?" So, I
16:27 think that that was really surprising to
16:28 me at least.
16:29 Well, I think it makes the robot look
16:31 purposeful. you know, it helps it make
16:34 it look like a a piece of equipment that
16:36 is used for doing something, you know,
16:39 and I I think that if you just have a
16:41 humanoid, there's a little less to
16:45 work with, you know, by that I mean a
16:47 literal humanoid, you know, it it it
16:50 starts just becoming a mannequin.
16:52 Yeah.
16:52 And you're just working toward the same
16:54 problem that everybody else is working
16:56 toward. And you know we were really able
16:58 to focus on you know solving our
17:01 customers problems and that's that's
17:03 ultimately you know what is driving the
17:06 shape of the robot and everything you
17:08 see you know what are those tasks that
17:10 the robot needs to do and you know what
17:14 is the most efficient way of doing them.
17:17 So when we talk about ROM and that's
17:20 range of motion
17:22 that's really uh every as much as the
17:26 robot can move what is it able to do
17:28 with its arms and legs where can we put
17:31 those grippers in space right so um you
17:35 know if you if you imagine making a snow
17:38 angel lying in the snow and waving your
17:41 arms and making as large wings that as
17:44 you can your arms are going to hit your
17:46 head at one point and they're going to
17:48 hit your hips at one point and that's
17:50 your your maximum range of motion in
17:52 that plane. So, we're that's that is
17:55 what we're doing for the robot. We want
17:57 to increase the range of motion as much
17:59 as possible. And maybe we can move the
18:01 head out of the way a little bit or move
18:02 the hips out of the way a little bit.
18:04 That's where, you know,
18:08 that's where the the what we're able to
18:10 do mechanically sort of crosses over
18:13 into what we can do with the behavior.
18:15 There's another piece to that as well
18:17 and that is the the fields of view for
18:19 the cameras. So, you know, both of these
18:22 things are kind of invisible geometry
18:24 that is around the robot. And if we we
18:27 think of the fields of view the same
18:29 way, we're we're projecting kind of a a
18:33 rectangle out from the cameras that gets
18:35 bigger as it gets farther away from the
18:37 robot. And that is so that the robot can
18:41 see its environment and so that it can
18:43 see its grippers and you know the other
18:46 parts of the robot and know where it is
18:48 in space. And with you know both of
18:50 these things there's a a safety
18:52 consideration as well where with the
18:56 cameras we want to make sure that the
18:58 robot can see people in the environment.
19:00 With the ROM, we want to make sure that,
19:02 you know, people don't become injured,
19:05 you know, and uh make sure that we have
19:08 uh adequate clearance and pinch points.
19:11 So, when we're thinking about the range
19:12 of motion and the fields of view of the
19:15 cameras, you know, that's something that
19:17 is going to directly impact the shape of
19:20 the robot.
19:21 Well, and that's essentially why or one
19:23 of the reasons why the cameras ended up
19:24 in the head.
19:25 Yeah. Yeah.
19:26 Putting them up there gives you a kind
19:27 of a fighting chance of being able to
19:29 see without the body including the view.
19:31 Yeah. If a human is working in a work
19:34 cell, they obviously can't see most of
19:36 what's around them. And that's just
19:38 accepted fact, right? If someone takes a
19:41 step backwards and bumps into something
19:43 like, well, that happens. But with a
19:45 robot, it's almost unacceptable, right?
19:48 Because technology should allow you to
19:50 avoid that situation. So, you know, we
19:54 employ these cameras all around the
19:56 head, but then it really dictates not
19:59 just where the cameras go, but what can
20:01 live around them, right? If it truly has
20:04 360 degrees of visibility, can operate
20:07 much more freely than a human can
20:10 because it's totally aware of what's
20:12 going on around it.
20:13 Yeah. And I think an example where the
20:17 robot is oluding the cameras sometimes
20:19 are these handles that we've built out
20:21 on the back of the robot and we've put
20:24 those there as a safety measure so
20:27 people can manipulate the robot easily
20:29 without putting their their hands in
20:31 harm's way. So these are two areas where
20:36 you know safety systems are competing
20:38 with one another. you know, the robot
20:40 can kind of look around and change its
20:42 position to see what's around those
20:45 handles. But, you know, these are
20:46 examples of competing concerns.
20:49 We initially try to
20:52 design something with no compromises,
20:55 right? But that quickly becomes
20:57 impossible and then you need to start
20:59 trading off features.
21:02 And I remember this interesting
21:04 conversation around these handles and
21:05 how yes it does olude specific angles of
21:10 view but it's really important that the
21:12 operators who might be handling you know
21:14 robot in the unpowered state have
21:16 something safe to grab on. So we're
21:18 going to figure out how to deal with
21:20 those blind spots through you know
21:22 intelligence and behavior etc. rather
21:25 than just you know pure hardware design.
21:28 Yeah, that's always really difficult to
21:31 negotiate
21:33 because I you know there are many
21:35 stakeholders and you have to
21:38 you have to talk to everybody and
21:40 nobody's going to get everything that
21:41 they want especially this early because
21:47 we'd like to believe that with you know
21:50 the modern reinforcement learning that
21:53 anything is possible
21:56 but it's going to take some time to
21:58 develop and we're sitting here in 2026
22:02 trying to design a piece of hardware.
22:04 We're trying to project forward what
22:07 behaviors are going to be possible in
22:08 2028 and 2030
22:11 and we need to know which ones we can
22:13 bet on and which ones we need to be a
22:16 bit more conservative about because if
22:18 we bet on a particular behavior that
22:20 proves harder to implement, well then
22:23 there's going to be a performance
22:25 regression we didn't intend. And I think
22:27 that's where the modularity of this
22:30 design is really going to help us,
22:32 right? Because we're not going to get it
22:34 100% right. And hopefully the modularity
22:38 will allow us to swap out parts of the
22:40 robot, redesign pieces easily, and kind
22:43 do that way faster than we would be
22:45 able to do it if we had a much more
22:46 integrated design.
22:48 Another great thing about modularity is
22:50 allows us to take a phased approach to
22:53 the hardware development. Like right
22:55 now, one of the big goals of the program
22:57 is giving the behaviors team a platform
22:59 which they can develop on. And whereas
23:02 we've got lots of hardware challenges we
23:04 need to solve, they don't all need to be
23:06 solved today. And so we continue to
23:09 iterate on those subsystems in parallel
23:13 [snorts]
23:13 and intercept them with the program when
23:15 they're ready such that as we get to the
23:18 key program milestones,
23:21 you know, we have the features we need
23:23 at that time, but developing those
23:25 features doesn't slow down the work that
23:27 has to happen. Now, installations
23:31 at our partner sites will allow us to
23:33 see in the real world, how does this
23:36 robot perform? What does it need more
23:38 of? What can it do with less of and
23:40 evolve that over the next couple of
23:42 hardware iterations?
23:44 Aaron, can you tell us a little bit
23:45 about how the head design, industrial
23:48 design evolved,
23:50 um, and why it is kind of the way it is
23:51 today? Yeah, when we were developing the
23:54 prototype, that's when we first put a
23:57 round head on the robot. That was a much
24:00 more humanoid,
24:03 you know, visual design.
24:04 The prototype.
24:06 The prototype. Yeah. Yeah.
24:08 Yeah.
24:08 So, you know, over over time, you know,
24:12 I got used to it and, you know, I think
24:14 that's really key with all of this robot
24:16 design. you know, you do something
24:18 that's um weird looking and you really
24:21 need to give it time to gestate, right?
24:24 You know, the difference between
24:25 prototype and product is the the
24:28 prototype is really just designed for
24:30 performance, right? All of those panels,
24:32 everything on the robot is just intended
24:36 to be functional. You know, we're trying
24:38 to make it look good and intentional.
24:41 So, when we looked at the product
24:44 version of the head, um, you know, there
24:47 were I think there were a lot of things
24:49 that I wanted to do just having thought
24:51 about it for a while. And one of them
24:53 was to create that that big sort of
24:56 silicone ring and think of that as, you
24:59 know, a light up icon. In addition to
25:02 that, it would be padding, right? So,
25:04 it's a big thick silicone ring. It can
25:06 light up and it can bump into things,
25:08 right? That presents a really nice
25:11 opportunity for us to explore the UX of
25:14 the robot. You know, when somebody walks
25:17 into the environment, what does the
25:18 robot do? What do people see? Does the
25:21 robot regard them? And and how does that
25:23 manifest itself in that light? You know,
25:25 I think that there's also something
25:26 really nice about that light ring. And,
25:30 you know, it avoids us having a face,
25:34 right? It's just a ring. It's at the top
25:36 of the robot. It's kind of a face
25:39 without, you know, becoming too literal.
25:41 You know, we don't have two eyeballs in
25:44 the robot. It definitely houses the
25:46 cameras and it and it looks around like
25:48 a real head, but uh you know, we don't
25:51 need to interact with the robot like we
25:53 would interact with a person.
25:55 Well, we have a rear [clears throat]
25:57 light ring also, right? And then the
25:58 little kind of antenna post. What What
26:01 are those about?
26:02 I think there's something really nice
26:03 about how the front and rear light ring
26:05 and that mast light came together.
26:08 you know, there's something iconic. You
26:11 know, it's it's it's a nice showcase for
26:14 the UX that we're going to be exploring
26:16 on the robot. Uh, another interesting UX
26:19 feature
26:21 uh was the neck pitch degree of freedom.
26:23 I know we debated that a lot whether or
26:25 not we absolutely needed it or should we
26:27 try to simplify the robot even more and
26:29 get rid of neck pitch, but we kind of
26:31 determined it was important for
26:35 potentially interacting with people.
26:37 Yeah, it's a it's an extra bit of
26:40 movement, right? There's another
26:42 actuator just to handle
26:45 a nod, right? So, in the beginning,
26:50 yeah, I think we wanted to have some way
26:53 to acknowledge when somebody walked into
26:55 the into the space, maybe gave a
26:58 command, um, you know, just a little
27:01 10°ree nod. And I think we went through
27:04 a period of of wondering whether we even
27:07 need that. And it turned out we we did
27:09 need it for perception. So, we needed it
27:12 to be able to see the robot's feet. you
27:14 know, we want to be able to see as close
27:17 to the robot as possible. And in order
27:19 to do that, just, you know, those few
27:20 degrees looking down really helped,
27:23 right?
27:23 And I think what we use even more is is
27:26 looking up. So, if we're reaching
27:27 towards something on a on a shelf, we
27:30 want to lift the the head up as much as
27:32 possible.
27:33 So, James, uh, what have you found to be
27:36 some of the more challenging aspects of
27:38 the head design from a comput and
27:40 sensing perspective? I lead the compute
27:42 and sensing team here, which is a
27:44 relatively new team. Um, [snorts] a team
27:47 that's increasingly important as we rely
27:50 on the robot to just process that much
27:53 more information. Right? In this this
27:56 era of AI models and reinforcement
27:59 learning, it'll be a long time before
28:01 our uh our friends in software say they
28:05 have more than enough compute power. So,
28:09 we're being asked to execute something
28:11 that's, you know, extremely efficient.
28:15 Um, I've worked on over 30 different
28:19 personal computers in my time at Apple.
28:22 And this is by far the most challenging
28:26 computer I've ever worked on. By far the
28:29 coolest computer I've ever worked on.
28:31 Um, and I say computer because the head
28:33 is just a computer on a neck. But I say
28:37 just um, but you know it has to have the
28:42 performance
28:43 of a powerful desktop computer
28:48 but the robustness
28:50 of like a mobile device. It needs to be
28:55 waterproof to survive certain
28:58 environments. Creating a very performing
29:01 computer that's also waterproof is a
29:03 really tough challenge and it's one that
29:05 we're continuing to iterate on. You
29:08 know, this is the fifth iteration of the
29:10 humanoid head at BD and there'll
29:14 [snorts] certainly be six and seven
29:16 before we ship this product because
29:18 there's still a lot of learning to do.
29:22 Have you worked on any computer that has
29:24 to move around and potentially get
29:26 bumped into things?
29:27 Well, you know, I've worked on lots of
29:31 MacBook Pros, MacBook Airs, and those
29:34 are handled, some of them, you know, not
29:37 very gently by our customers. Uh it's
29:40 very challenging to make those robust,
29:43 but
29:45 the types of events that those are
29:48 expected to survive are nothing compared
29:52 to, you know, a humanoid robot um
29:56 potentially, you know, tripping and
29:59 falling from 2 meters in height and
30:01 impacting the edge of a table. You know,
30:03 these are all real scenarios during
30:07 these early stages of humanoid
30:09 development, right? Most companies show
30:11 lots of videos of robots never falling.
30:14 The reality is in the lab, robots fall
30:17 all the time. And if the head broke
30:20 every time the robot fell, it would be a
30:22 disaster.
30:24 And so,
30:26 we need something that's really, really
30:28 robust. Um, and over time we have to
30:30 figure out exactly how robust it needs
30:32 to be to survive
30:36 the real world demands on a robot
30:38 because you know anytime you make
30:39 something robust, you're probably
30:42 trading something else. You're probably
30:44 trading lightness or you're trading cost
30:47 or you're trading assembly complexity.
30:50 [snorts] And so we need to be sure we
30:53 are not overengineering this. And in
30:56 order to know that we have to really
30:59 think hard about what is the robot
31:02 environment in steady state. What are
31:06 the expectations of the customer on what
31:09 a robot can survive?
31:12 Um, and you know, the unique thing about
31:15 a robot is if it does take damage,
31:19 this robot's designed to be repaired and
31:22 uh,
31:24 brought back online in the matter of
31:26 minutes. So maybe you build that into
31:29 your strategy. You take advantage of the
31:32 fact you can repair it quickly in the
31:34 calculation of how robust you need to
31:36 make this. So, it's a very challenging
31:40 problem to solve because there is no
31:42 right answer. There are lots of
31:45 potential answers and we need to decide
31:48 on the philosophy that we're going to
31:50 employ in the design of this. Well, this
31:53 was fun. Uh, thanks for joining us and
31:55 visit bostonynamics.com if you want to
31:58 learn more.