Latest / Automated with Brian Heater / Zachary Jackowski on Generalization in Robotics and the Reality of Deploying Robots in the Real World

Transcript

Zachary Jackowski on Generalization in Robotics and the Reality of Deploying Robots in the Real World · Automated with Brian Heater

• 0:00
  SPEAKER_00: One of the things I enjoy most about our team here is we understand designing robots well enough to understand when we want to conform to like a natural form or an industry standard.
• 0:14
  SPEAKER_00: And we're totally comfortable with just going off and doing our own thing.
• 0:18
  SPEAKER_00: Like we actually understand the fundamental constraints on the engineering, the fundamental physics here.
• 0:23
  SPEAKER_00: And the team is just completely unbounded in going in a new direction because there's so much understanding and confidence about the fundamentals of the design problem.
• 1:02
  SPEAKER_01: I learned a lot about the company's approach to humanoids from the chat and how Bots Dynamics is approaching research platforms after shifting much of his efforts towards commercialization.
• 1:14
  SPEAKER_01: If you like the show, don't forget to like and subscribe.
• 1:17
  SPEAKER_01: Please check out our newsletter over at automated.fm.
• 1:21
  SPEAKER_01: Also, automated is coming to Boston or taking our first ever road trip in a few weeks.
• 1:27
  SPEAKER_01: We are recording a bunch of episodes in person, including some in front of a live audience.
• 1:32
  SPEAKER_01: We are also hosting a happy hour from 4 to 6 p.m.
• 1:35
  SPEAKER_01: at Mass Robotics on April 8th.
• 1:38
  SPEAKER_01: And thank you to Maxon for sponsoring that event.
• 1:41
  SPEAKER_01: And with that, please enjoy this conversation with Zachary Jekowski from Bosta Dynamics.
• 1:48
  SPEAKER_01: I figured a good place to start would be to reflect on uh Robert Plater's time at Bosta Dynamics.
• 1:55
  SPEAKER_01: Obviously, like the history of Bosta Dynamics is Robert Plater's time at Boston Dynamics, but you know, a really big role that he has played as CEO is making that transition into commercialization.
• 2:05
  SPEAKER_01: I mean, that's kind of essentially what he took the throne from Mark to do over the last few years.
• 2:11
  SPEAKER_00: It's still kind of sinking in a little bit that we're gonna have to figure out how to do this whole thing without Rob by our side at every step.
• 2:20
  SPEAKER_00: It was Rob almost 15 years ago now who gave me my offer letter to Boston Dynamics.
• 2:26
  SPEAKER_00: I've been working directly on his team for such a long time now.
• 2:30
  SPEAKER_00: It's hard to express kind of how much of kind of what you've seen about how the company has excelled both on the product side and the commercial side is due to Rob's inexhaustible energy and perseverance.
• 2:47
  SPEAKER_00: The number of major corporate events moving from an independent company to Google to SoftBank to Hyundai, the fact that those all went well, it was not like a thing that just happened.
• 2:59
  SPEAKER_00: That was Rob's perseverance, you know, just working day in and day out to make those transitions work and negotiating all the details.
• 3:09
  SPEAKER_00: Because of that work, we are very much on our feet now.
• 3:12
  SPEAKER_00: The robotics product design research organization and on the commercial side, things are just cranking along.
• 3:20
  SPEAKER_00: That's what gave him the comfort to say, this is my opportunity.
• 3:24
  SPEAKER_00: Like, yeah, you guys take it from here.
• 3:26
  SPEAKER_01: I was matching up the timeline.
• 3:27
  SPEAKER_01: So um, you know, 15 years ago, so you started, you you came on um 2011, um, and you came on from really like pure research.
• 3:36
  SPEAKER_01: You were over at MITC sale.
• 3:38
  SPEAKER_01: You come on, and then two years later, Google buys Boston Dynamics, and then four years later, Softbank, and now it's Hyundai.
• 3:46
  SPEAKER_01: So obviously the company has has been through a lot in the 15 years since since you've been there.
• 3:51
  SPEAKER_01: I mean, it must have been your first two years at that company, it must have been very interesting, really, undergoing that period of being again purely research and and in that sense to a certain degree, very similar, I assume, to what you were doing over at CSale under, I think, I guess Russ T.
• 4:07
  SPEAKER_01: Drake when he was over there, uh, to again, this big corporation under Google, uh, an investment bank under SoftBank, and now uh commercial arm.
• 4:17
  SPEAKER_00: There's a few things to tease apart there, though.
• 4:19
  SPEAKER_00: Yeah.
• 4:19
  SPEAKER_00: I've always been pretty entrepreneurial and product oriented at heart.
• 4:24
  SPEAKER_00: Even during undergrad and grad school at MIT, I had a startup at the time.
• 4:28
  SPEAKER_00: And so like I I've always had a passion for doing robotics, but also a passion for doing product.
• 4:35
  SPEAKER_00: At Boston Dynamics, over the first like four years that I was here, the big thing that was happening was it was becoming increasingly clear that the research that we were working on, that it was working.
• 4:50
  SPEAKER_00: And that if we wanted to push it to the next level and do you know bigger and better things and more impactful research, it was clear that the next thing we had to do in the pursuit of the frontier of robotics was go into product and be more commercial.
• 5:07
  SPEAKER_00: If you look at it from the perspective of the bunch of people at Boston Dynamics came here to work on the forefront of robotics and to make an impact in our work, it makes sense that that transition was not as big as you would have thought it was.
• 5:24
  SPEAKER_00: Everyone's excited about being on the forefront and doing bigger and better things.
• 5:28
  SPEAKER_00: And everyone was mentally acclimated to the idea that the uh blue sky research part, you know, we kind of tapped that out.
• 5:36
  SPEAKER_00: It's like it's time to figure out how to go bigger and better now.
• 5:39
  SPEAKER_00: Obviously, you know, the transition from a 65-person, fully independent private company to learning how to live at Google was interesting.
• 5:48
  SPEAKER_00: But on the other hand, we've kind of always been Boston dynamics, and we are to this day.
• 5:54
  SPEAKER_00: And the company itself has a very supportive, very tight-net engineering culture and now commercial and product culture.
• 6:03
  SPEAKER_00: It's a bigger than it was family, but it's still very much a family here.
• 6:08
  SPEAKER_00: You know, it's everyone in now a few buildings now, but it's a cohesive whole, and it's always felt like a cohesive whole.
• 6:15
  SPEAKER_00: And we learned how to be a cohesive whole inside Google and work with Google.
• 6:20
  SPEAKER_00: You know, we learned how to work with SoftBank, and now we're learning how to do the same thing with Hyundai, where, you know, it's still Boston dynamics.
• 6:29
  SPEAKER_00: We all have Boston Dynamics badges, you know, everything that goes along with that.
• 6:33
  SPEAKER_00: But there is this big entity that knows how to do things that we don't know how to do and has access to resources that we don't have on our own.
• 6:41
  SPEAKER_00: You know, it's still comfortable, it's still Boston dynamics and it's still working.
• 6:45
  SPEAKER_01: That's really interesting because one of the things that you and I spoke about the last time we spoke, maybe like a month or two back, was that the original version of the electric atlas that we saw, I think R1 is what you were calling it in-house.
• 6:56
  SPEAKER_01: Yep.
• 6:56
  SPEAKER_01: That it was designed to be a research robot.
• 6:58
  SPEAKER_01: And in that sense, it was similar to the original hydraulic atlas or the earlier robots that we saw.
• 7:05
  SPEAKER_01: In that sense, and insofar as Basa Dynamics does continue to be a research um lab and does do these kind of you know boundary-pushing robots.
• 7:15
  SPEAKER_01: Is there still somewhat of a distinction between kind of the research side and the commercial side?
• 7:20
  SPEAKER_00: Up until last year, I would have said yes, that each of our products follows this arc where it's first in research, then the research is kind of working, and then you know, we're going full commercialization.
• 7:33
  SPEAKER_00: We've got spot, stretch, and atlas.
• 7:35
  SPEAKER_00: And so, you know, spot was first and it was on that arc, and then stretch started, and it was on that arc.
• 7:40
  SPEAKER_00: And now Atlas is on that arc.
• 7:42
  SPEAKER_00: Even though we're focused on product and we're driving there really hard, and we're designing this organization like a commercial entity, it is still very much research.
• 7:53
  SPEAKER_00: It's just now it's all centered around VLA design and agentic VLM behavior and how we're gonna do dexterous manipulation.
• 8:01
  SPEAKER_00: It's part of a bigger product story, but it's still very much research.
• 8:06
  SPEAKER_00: The size of the research team on Atlas is as big as Boston Dynamics in its entirety was when I joined the company.
• 8:14
  SPEAKER_01: This is something I was trying to figure out, and I and it's always hard to tell, you know, as somebody who obviously like we do a podcast and we try to figure out um, we try to figure out names for podcast episodes that will hook you in, but are also are like technically accurate.
• 8:28
  SPEAKER_01: And I was looking at the video of um obviously like the very impressive video of the R1 doing the flip and everything else, and and there was something about the title that really made it feel like this was kind of the R1's last hurrah.
• 8:43
  SPEAKER_01: Is that right?
• 8:44
  SPEAKER_01: I mean, are you effectively retiring the research version of the robot or will it still be around in some form to continue to kind of push that side of things?
• 8:53
  SPEAKER_00: It's really done.
• 8:54
  SPEAKER_00: So we got back from CES.
• 8:56
  SPEAKER_00: We had a bunch of working robots from CES, and our team did a just a big exhale, letting out all that stress and pressure.
• 9:05
  SPEAKER_00: And then we decided, hey, let's go do some really fun stuff uh with these robots for a couple weeks.
• 9:12
  SPEAKER_00: But kind of behind that is the new version of the robot is so much better on every axis.
• 9:19
  SPEAKER_00: So, yes, the R1 robots look a little bit more human, and they have a couple specs where they have some really special performance, but honestly, the new robots are so much more reliable, predictable, repairable.
• 9:35
  SPEAKER_00: The sensing and compute aspects of them make so much more sense.
• 9:39
  SPEAKER_00: We were ready to let the R1 robots go.
• 9:42
  SPEAKER_00: We're making static display models.
• 9:44
  SPEAKER_00: We've got one up in our museum now.
• 9:46
  SPEAKER_00: Yeah, we're putting them away.
• 9:47
  SPEAKER_00: They're on a code branch.
• 9:48
  SPEAKER_00: There's a couple people still using one, but that branch is getting increasingly uh divergent from trunk.
• 9:56
  SPEAKER_01: I got the sense, you know, the last time we spoke that kind of the idea behind R1 and the idea behind starting as a research robot first was, you know, let's build a robot.
• 10:05
  SPEAKER_01: Price isn't a factor, you know, weight isn't a factor.
• 10:08
  SPEAKER_01: Certain safety concerns as far as like having to like work alongside people aren't a factor in the same way that this will be, you know, when we deploy these things.
• 10:15
  SPEAKER_01: Therefore, this is going to be like bleeding edge state of the art.
• 10:19
  SPEAKER_01: Now we're moving towards productization, you know, obviously, like weight, pricing, repairability, scalability are all concerns.
• 10:26
  SPEAKER_01: Don't you necessarily have to make some compromises and don't you necessarily sort of lose some of that functionality in the process?
• 10:33
  SPEAKER_00: It's a little bit more complex than that.
• 10:35
  SPEAKER_00: Um, you know, first of all, um, we build research machines in or really as a a point of focus.
• 10:44
  SPEAKER_00: Like if you launched from day one and just said, oh, we want to build the whole humanoid robot product enchilada, there are so many decisions and technical threads to pull on and information you fundamentally do not have yet about what makes the robot work well, um, that you just get bound up.
• 11:04
  SPEAKER_00: And so, you know, part of the purpose of making a research robot is just an artificial simplification of the problem so that you can go gather the information necessary to design the whole thing.
• 11:16
  SPEAKER_00: But the second aspect is you know, now that we designed that R1 robot and we understand how all the pieces fit together, you know, when we built that robot, you know, we didn't yet understand how all the pieces fit together precisely.
• 11:29
  SPEAKER_00: So, yes, you can spend a lot of money and buy really nice materials and you know, kind of push the limits of performance in a lot of ways.
• 11:38
  SPEAKER_00: But in a lot of ways, when you build a first generation robot like that, it's an it's an unbalanced machine.
• 11:45
  SPEAKER_00: The fact that you went to the extremes on some parts, they don't all play together perfectly well.
• 11:50
  SPEAKER_00: And so when you go and you do a second generation design, you now know, hey, you know, if I make this, I can make this a little bit weaker, I can make this a little bit heavier.
• 12:00
  SPEAKER_00: I didn't need that whole thing.
• 12:02
  SPEAKER_00: So like the R1 robot has three computers in it: two x86 computers and a Jetson in the head.
• 12:09
  SPEAKER_00: The new robot has a single Thor, and that's an enormous simplification and actually an enormous jump in compute power.
• 12:17
  SPEAKER_00: And we get that in a lot of places in the robot.
• 12:20
  SPEAKER_00: So not just the compute, but the actuator designs.
• 12:24
  SPEAKER_00: We know so much more about how to build super performant actuators.
• 12:28
  SPEAKER_00: And we actually know how to build super performant low-cost actuators because we figured some things out that let us have our cake and eat it too.
• 12:37
  SPEAKER_00: When you design a product, it's not always just make it worse to make it cost less.
• 12:43
  SPEAKER_00: There are plenty of places where you can legitimately be smarter or use information that you didn't have the first time to make it fundamentally better in every access.
• 12:55
  SPEAKER_01: Yeah, if I'm remembering correctly, was it two actuators on the entire body or and three if you count the hand?
• 13:01
  SPEAKER_01: Yeah, that's exactly right.
• 13:03
  SPEAKER_01: The thing that really impressed everybody in the first video of the R1 was it's standing up, it could change direction, and that's a sign of incredibly powerful actuators, right?
• 13:12
  SPEAKER_01: That it could flip on the spot.
• 13:13
  SPEAKER_01: Were you able to uh maintain that kind of uh power and adaptability while paring it down to this many actuators?
• 13:22
  SPEAKER_00: Yes, in short, the actuators are phenomenal.
• 13:25
  SPEAKER_00: In fact, it's actually the opposite direction of that.
• 13:28
  SPEAKER_00: So the one piece about the production version of the robot that leaves us a little queasy still about the two actuator decision is these giant elbows.
• 13:38
  SPEAKER_00: Uh, there are actually more places on the robot where it's has excess strength and excess power capability than places where it's actually underpowered.
• 13:49
  SPEAKER_00: Yeah, you saw like the specs on like how much the production robot can lift and such.
• 13:54
  SPEAKER_00: In some ways, that's a consequence of our simplification work.
• 13:58
  SPEAKER_00: We got so good at making these actuators really strong that we were able to simplify the robot down to mostly the stronger version of the actuators.
• 14:08
  SPEAKER_01: Yeah, so I think you pointed out too, and and and this was partially a product of my, you know, unfortunately not having been at CES this year and seen it in person, was you were you were surprised by the reaction in that you thought people's reaction would be that it was more of a radical departure design-wise.
• 14:27
  SPEAKER_01: And you pointed to the legs, because the legs are a really big design change.
• 14:31
  SPEAKER_01: And why necessarily did you have to make such a big design change to the legs?
• 14:36
  SPEAKER_00: Okay, so there's a couple things in there again.
• 14:38
  SPEAKER_00: The biggest driver for that design is actually making the robot safer.
• 14:44
  SPEAKER_00: One of our top goals with the production version of the robot was to drive out as many pinch points as possible.
• 14:51
  SPEAKER_00: If you look at most robots, you like look behind the knee.
• 14:54
  SPEAKER_00: There's a pinch point there.
• 14:56
  SPEAKER_01: Literally somewhere where like a finger can get caught.
• 14:58
  SPEAKER_00: Yeah, exactly.
• 14:59
  SPEAKER_00: Some of the worst examples even have like exposed linkages and stuff.
• 15:02
  SPEAKER_00: But like anywhere that you can get like a finger in and you know crunch someone, we really don't want that kind of situation going on in the robot.
• 15:11
  SPEAKER_00: And so that pushed us down this path of uh trying to make as many of the uh axes on the robot offset axes as possible.
• 15:21
  SPEAKER_00: That's a kind of a common design pattern in robotics.
• 15:24
  SPEAKER_00: You look around, you'll see that repeated in most industrial cobalt arms and such.
• 15:29
  SPEAKER_00: Second, you know, maximizing the range of motion capabilities.
• 15:32
  SPEAKER_00: You want these offset joints, otherwise, your joint interferes with itself.
• 15:36
  SPEAKER_00: Those are the kind of the two drivers.
• 15:38
  SPEAKER_00: One of the things I enjoy most about our team here is we understand designing robots well enough to understand when we want to conform to like a natural form or an industry standard, and we're totally comfortable with just going off and doing our own thing.
• 15:55
  SPEAKER_00: Like we actually understand the fundamental constraints on the engineering, the fundamental physics here.
• 16:01
  SPEAKER_00: And the team is just completely unbounded in going in a new direction because there's so much understanding and confidence about the fundamentals of the design problem.
• 16:11
  SPEAKER_01: We know what a lot of the pros of the humanoid form factor are.
• 16:15
  SPEAKER_01: We know what some of the drawbacks are.
• 16:17
  SPEAKER_01: But the way I've I sort of started framing it recently for me is the way I'm thinking about in 2026 is that I'm no longer thinking about the humanoid form factor as being uh an endpoint for robot design, but rather a starting point.
• 16:31
  SPEAKER_00: Aaron Powell That pretty much mirrors our philosophy too.
• 16:34
  SPEAKER_00: There are lots of places where you know we can see making wheeled platforms with arms, fixed arms, forearmed robots, you know, all sorts of things.
• 16:44
  SPEAKER_00: But we do have to pick a thing to do first.
• 16:46
  SPEAKER_00: Yeah.
• 16:47
  SPEAKER_00: In our judgment, it makes the most sense to do the humanoid first because it has such compelling generality.
• 16:56
  SPEAKER_00: And I really think that the theme of modern robotics that everyone's discovering is the importance of generality.
• 17:05
  SPEAKER_00: We're kind of being taught that by the success of LLMs and uh VLMs.
• 17:11
  SPEAKER_00: Okay, so it's cool that these things appear kind of intelligent, but the the big revelation is hey, we've been building specialized software tools.
• 17:21
  SPEAKER_00: Yeah.
• 17:21
  SPEAKER_00: And once we got shown a way to build a generalist tool that can do these specialized things, the lid just got blown off, like software engineering and software in general.
• 17:32
  SPEAKER_00: We're seeing the same thing in robotics, and that the magic is the generality.
• 17:36
  SPEAKER_00: And the magic is it's one instantiation that can do everything.
• 17:42
  SPEAKER_01: I mentioned your work at CSale before.
• 17:45
  SPEAKER_01: I think that the work maybe you were doing alongside Russ.
• 17:47
  SPEAKER_01: I know this is something that he had been teaching for a long time.
• 17:51
  SPEAKER_01: And I'm wondering if this dovetails, because I'm not super familiar with the underactuated robotics, but is that something that kind of informs the direction that you've been going in with Atlas?
• 18:04
  SPEAKER_00: Underactuated robotics is really just the term for the field of control system engineering where you don't have direct control over every degree of freedom in the system.
• 18:16
  SPEAKER_00: So like an industrial robot that's bolted to the ground, you know, you can apply direct torque control to every portion of that robot state vector.
• 18:26
  SPEAKER_00: With a balancing robot, a walking robot, you know, you can control the joints of the robot, but you don't have direct control over the base link of the robot, or like you know, your body is not attached to another arm holding it to the ground.
• 18:41
  SPEAKER_00: Like you have to influence the evolution of the dynamics of the robot through how you interact with the world.
• 18:48
  SPEAKER_00: And so like a foot can't pull on the world, it has a limited amount of torque it can exert on the world.
• 18:54
  SPEAKER_00: And so that's the definition of the underactuated term and how it relates to robotics.
• 18:59
  SPEAKER_00: Yeah, the things that we were working on at CSEL back in the day, we were focused on the problems of balance and manipulation and how you design control systems for uh you know robotics problems that you don't have full control of.
• 19:16
  SPEAKER_00: Yeah, that's kind of where we are today.
• 19:18
  SPEAKER_00: So, of course, that yeah, that thinking has influenced where we are.
• 19:21
  SPEAKER_00: Yeah, when I started my grad school work back in 2009 or eight, what folks would call the ice age of machine learning.
• 19:32
  SPEAKER_00: And Russ was one of the kind of true believers still in reinforcement learning.
• 19:37
  SPEAKER_00: Back when I joined the lab, he always had that saying that birds aren't you know learning how to solve Navia Stokes.
• 19:44
  SPEAKER_00: It's one of the equations of motion for fluid flow.
• 19:47
  SPEAKER_00: Yeah, model-free learning.
• 19:49
  SPEAKER_00: And so that that's what we were working on.
• 19:51
  SPEAKER_00: At the lab, we spent most of the time working on model-based control methods and optimization-based control, yeah, normal robotics stuff.
• 19:59
  SPEAKER_00: Finally, in the late 2010s, uh, the world kind of snapped back to the original design philosophy of that lab and what we were all working on initially.
• 20:10
  SPEAKER_01: Just yesterday, as a matter of fact, we did uh we had a really uh a great conversation with uh Sergey Levine.
• 20:17
  SPEAKER_01: You know, he's at uh obviously uh Berkeley, but also physical intelligence.
• 20:20
  SPEAKER_01: And we were talking a lot about data collection.
• 20:24
  SPEAKER_01: I use the phrase brute force to describe, you know, this idea of kind of just getting robots out in the world to collect as much data as possible.
• 20:31
  SPEAKER_01: And it strikes me now that if that's the method that we're going to use, if that's the really the way to create generalized robots, is to just get as much real world data as we can, then we're in this limital space right now, right?
• 20:44
  SPEAKER_01: It's like a chicken in the egg.
• 20:46
  SPEAKER_01: How do we get robots that are generalized enough that we can get them out in the world in order to collect that data that we need?
• 20:51
  SPEAKER_01: To a certain extent, I assume that the approach, and I do see Botswana Dynamics and other companies taking this, is it's sort of baby steps, right?
• 20:58
  SPEAKER_01: You get it out there, you get it doing sort of one job at a time, and then you start taking on more work.
• 21:05
  SPEAKER_00: That's a correct characterization of the situation.
• 21:08
  SPEAKER_00: I haven't heard it called a liminal space yet.
• 21:10
  SPEAKER_01: Oh, I I I coined it, so yeah, we can I'd say that's pretty accurate.
• 21:14
  SPEAKER_00: There is a really important thing that isn't super widely appreciated about the data though, which is we all have this hope and a fair expectation at this point that you get your data together and you're going to see generalization of behavior capability.
• 21:32
  SPEAKER_00: But I can tell you it's pretty firmly known now that if you get out there and you do all of one thing, like yeah, so we're showing this sequence automotive part sequencing behavior.
• 21:44
  SPEAKER_00: If we deploy robots and just do automotive part sequencing, we are not going to produce the kind of generalization that we're hoping for.
• 21:53
  SPEAKER_01: You mean like like you you you you do automotive part sequencing, it's not going to be able to do my dishes, something like that?
• 22:00
  SPEAKER_00: Exactly.
• 22:00
  SPEAKER_00: Like quality and the diversity of the data is every bit as important as the quantity of the data.
• 22:08
  SPEAKER_00: It's not as simple as, oh, just get good at one thing and get out there and collect lots of data.
• 22:14
  SPEAKER_00: Yeah, the smart folks are constantly pushing kind of a level of discomfort about doing new things, you know, new kinds of actions, new environments, rather than just kind of hill climbing up, getting really, really good at this one specific thing.
• 22:31
  SPEAKER_01: I'm kind of curious to get your approach to generalization from a research standpoint.
• 22:37
  SPEAKER_01: As far as there are companies out there that are saying we're going after the home first, right?
• 22:41
  SPEAKER_01: There are companies that are out there that are saying we're starting to sort of experiment with the home.
• 22:44
  SPEAKER_01: There are companies that are saying we can see use cases beyond industrial on the horizon.
• 22:51
  SPEAKER_01: Obviously, generalization is something that is like somewhere, you know, you would love to have a generalized, general purpose uh robot.
• 23:01
  SPEAKER_01: Um so what does that mean?
• 23:03
  SPEAKER_01: Does that mean that like that you look at the home?
• 23:06
  SPEAKER_01: Does that mean partnering and doing some sort of work and seeing how an atlas would react in a different environment?
• 23:13
  SPEAKER_00: Yeah, there's a few things buried in there.
• 23:15
  SPEAKER_00: Even in like industrial and commercial environments, you need generalization.
• 23:20
  SPEAKER_00: Even just limiting yourself to working in a car plant, yeah, you need generalization.
• 23:26
  SPEAKER_00: There are too many subtasks and weird little things.
• 23:30
  SPEAKER_00: One of our big messages, we're part of Hyundai, and Hyundai is so beautifully good at designing car plants.
• 23:38
  SPEAKER_00: And so, like if you go to a Hyundai car plant, there is a very good reason why every task in that plant that isn't automated is not automated.
• 23:48
  SPEAKER_00: Like they know this stuff backwards and forward so well.
• 23:51
  SPEAKER_00: And the reason is that there is such a diversity of behaviors around all of these things that aren't automated yet in the car plant.
• 24:00
  SPEAKER_00: And so you need that generalization to do that stuff.
• 24:03
  SPEAKER_00: There's plenty of diversity to engage with just in that context.
• 24:07
  SPEAKER_00: The thing you do get with that context is you get some regularity.
• 24:12
  SPEAKER_00: So there's a huge diversity, but you get to engage with that diversity in a repeated form.
• 24:19
  SPEAKER_00: And you get to work around trained people, trained adults who understand how to work around a piece of equipment.
• 24:28
  SPEAKER_00: You don't have those two things in the home.
• 24:30
  SPEAKER_00: And I think that's going to make the home obviously a place that we are going to get to, but it's a little bit more complex to engage with.
• 24:39
  SPEAKER_00: If you're trying to build something that's commercially successful pretty early, we're highly interested in home applications.
• 24:47
  SPEAKER_00: Uh, we're also just highly interested in kind of getting our commercial juices flowing, in deploying robots to customers that genuinely see a return on investment on deploying them.
• 25:00
  SPEAKER_00: That's what we're going to get with the industrial environments.
• 25:02
  SPEAKER_01: Yeah.
• 25:03
  SPEAKER_00: The opportunity in commercial service and home markets is huge.
• 25:08
  SPEAKER_00: And of course, we'll be playing in that area.
• 25:11
  SPEAKER_01: Maybe this is a better way to frame it, and hopefully with fewer words.
• 25:15
  SPEAKER_01: Um, we spoke about the research part.
• 25:18
  SPEAKER_01: You said that's a big part, and you said that's what this um the productized version of Atlas is being used for now.
• 25:26
  SPEAKER_01: To that extent, outside of selling to like a Hyundai or these systems that like you want to put out in the world for commercial reasons.
• 25:36
  SPEAKER_01: Does the current production atlas have a role to play in research being done outside of that environment?
• 25:42
  SPEAKER_00: Oh, yeah.
• 25:43
  SPEAKER_00: We'll definitely be doing our first experiments in commercial and service robotics with the production version of Atlas that you see today.
• 25:51
  SPEAKER_00: Yeah.
• 25:52
  SPEAKER_00: Yeah.
• 25:52
  SPEAKER_00: I'm sure we'll spend some time in home environments too, making beds and folding laundry and cleaning countertops and all that stuff too.
• 25:59
  SPEAKER_00: Amazing platform.
• 26:00
  SPEAKER_00: It's amazingly capable, probably a little bit too capable for a home environment.
• 26:05
  SPEAKER_00: That's a great thing if you're going and doing research.
• 26:07
  SPEAKER_01: What does it mean to be too capable?
• 26:09
  SPEAKER_00: I think the robot is slightly heavier than we would want for a home environment.
• 26:14
  SPEAKER_00: Yeah.
• 26:14
  SPEAKER_00: In these industrial applications, you have the luxury of having the people that you are around being a known thing.
• 26:21
  SPEAKER_00: So you are around able-bodied adults who are trained on interacting with your robot in a home environment.
• 26:29
  SPEAKER_00: You do not have that.
• 26:31
  SPEAKER_00: You've got kids running around, you've got pets, and I think that's going to lend itself to a pretty different kind of platform once you get to like the mass scale.
• 26:39
  SPEAKER_00: But again, great for experimenting and research.
• 26:42
  SPEAKER_01: Sounds like a lot of these things are just sort of come down to bandwidth, right?
• 26:47
  SPEAKER_01: I mean, we're using words like generalization.
• 26:50
  SPEAKER_01: What do we prioritize?
• 26:51
  SPEAKER_01: What can we take on?
• 26:52
  SPEAKER_01: You have a lot of resources, obviously.
• 26:54
  SPEAKER_01: Hyundai has a lot of resources, but you still need to prioritize things.
• 26:57
  SPEAKER_01: Yep.
• 26:57
  SPEAKER_01: So what do you put first?
• 26:58
  SPEAKER_01: Just because Atlas has legs doesn't mean you believe that there aren't going to be applications where a humanoid with wheels might not make sense.
• 27:07
  SPEAKER_00: Yeah, that's completely fair.
• 27:08
  SPEAKER_00: This is really about focus.
• 27:10
  SPEAKER_00: No matter how many resources you have to apply, whether that's your number of people or number of dollars, for a project like this, the fundamental constraint you have is how much focus your researchers and leaders can apply to directing those resources.
• 27:29
  SPEAKER_00: You fundamentally have to have people at the top that are paying attention to the details.
• 27:34
  SPEAKER_00: Um, that's kind of where you run out of bandwidth.
• 27:37
  SPEAKER_01: One of the things I'm curious too, because you you obviously spent uh quite a bit of time on the SPOT team.
• 27:42
  SPEAKER_01: I think you were like heading up the SWOT team as well.
• 27:45
  SPEAKER_01: Yep.
• 27:45
  SPEAKER_01: One of the innovations there, um, especially towards the end, was this idea of turning SPO into a developer platform, right?
• 27:53
  SPEAKER_01: Like let's do iPhone Android version of a robot where we don't have to develop all the applications.
• 27:59
  SPEAKER_01: Yeah.
• 27:59
  SPEAKER_01: Is that something that continues to make sense even as we're talking about VLAs and VLMs, is to sort of open up these applications to third parties?
• 28:08
  SPEAKER_00: So Spot actually launched as a developer platform, and the evolution of it was us kind of figuring out that just throwing a platform out there isn't enough to make a product successful.
• 28:23
  SPEAKER_00: It turns out that that's really hard work.
• 28:25
  SPEAKER_00: If there's one thing that other people aren't going to do for you, it's really hard work.
• 28:31
  unknown: Yeah.
• 28:31
  SPEAKER_00: And so, yeah, we put the developer platform out there.
• 28:34
  SPEAKER_00: It was exciting.
• 28:35
  SPEAKER_00: We had super good engagement from a lot of partners.
• 28:39
  SPEAKER_00: But fundamentally, you know, we needed a team of very focused, very well resourced roboticists in order to get promising applications over the line in terms of feature completeness and reliability.
• 28:52
  SPEAKER_00: We kept the developer platform aspect, but we picked some primary applications that we made successful ourselves.
• 28:59
  SPEAKER_00: With the humanoids, we're going to cut to the chase a little bit.
• 29:02
  SPEAKER_00: Having done that story, we're picking manufacturing to be our successful application.
• 29:08
  SPEAKER_00: We will certainly engage partner companies and such.
• 29:12
  SPEAKER_00: We'll certainly have your research efforts inside.
• 29:15
  SPEAKER_00: Um, all those will benefit from platform aspects.
• 29:18
  SPEAKER_00: We'll do those, and you're totally right that the evolution of the technology stack toward VLM-based systems that are fundamentally kind of prompted rather than painstakingly coded makes that a lot easier.
• 29:32
  SPEAKER_00: It does push the work of like making that system reliable back onto the training side now.
• 29:39
  SPEAKER_00: So just because it's easier to interact with that system and fine-tuning it by you know prompting it differently for different situations doesn't get around the fact that, hey, when you find real problems with a system, you need to understand them, root cause them, and go retrain the system for proper behavior.
• 29:57
  SPEAKER_00: And so that's still going to be a limitation.
• 29:59
  SPEAKER_01: It sounds like for you and the team that uh to a certain extent you were really able to hit the ground running with this Atlas project because of the work that you did on spot.
• 30:09
  SPEAKER_01: Exactly.
• 30:09
  SPEAKER_01: When did you personally make the transition between teams?
• 30:13
  SPEAKER_00: Oh, I'm kind of losing track of time.
• 30:16
  SPEAKER_00: Yeah, that's fine.
• 30:18
  SPEAKER_00: Maybe about two years ago.
• 30:20
  SPEAKER_00: Okay.
• 30:20
  SPEAKER_00: So fairly recently.
• 30:22
  SPEAKER_00: When did work on the R1 start?
• 30:24
  SPEAKER_00: It depends on what you call the iterations.
• 30:27
  SPEAKER_01: So when did you start working on the electric version?
• 30:29
  SPEAKER_01: Let's say that.
• 30:30
  SPEAKER_00: Two and a half years before I started on Atlas.
• 30:33
  SPEAKER_01: It's funny because I was also looking at the timeline from this standpoint, your timeline from this standpoint.
• 30:37
  SPEAKER_01: So 2011, and then I guess officially Atlas, hydraulic atlas is announced about two years after you had started, right?
• 30:46
  SPEAKER_01: Yes.
• 30:46
  SPEAKER_01: Was that a difficult transition at all from the spot team to the Atlas team?
• 30:50
  SPEAKER_01: I mean, how dramatic of a difference is that?
• 30:53
  SPEAKER_00: Like I I mentioned, you know, right at the start about just kind of like how Boston dynamics feels.
• 30:58
  SPEAKER_00: The spot team, the Atlas team, and the stretch team, we're all like, you know, literally down the hall from each other.
• 31:05
  SPEAKER_00: We all have lunch together and we all build on a common tech stack.
• 31:09
  SPEAKER_00: And so even though like the teams operate, you know, relatively independently so that they can choose the fastest path to success, fundamentally they're all pretty similar.
• 31:19
  SPEAKER_00: Moving between them is pretty fluid.
• 31:21
  SPEAKER_00: It's a you know, all of them are very familiar spaces to folks inside Boston Dynamics.
• 31:26
  SPEAKER_01: Stress doesn't get enough love.
• 31:27
  SPEAKER_01: No, it doesn't.
• 31:28
  SPEAKER_01: What are some of the learnings on that side that kind of have informed the way Boston Dynamics approaches robots?
• 31:34
  SPEAKER_00: One of the lessons is if you want to move boxes at super high rates, super predictably, even for the company that you know kind of pioneered legged robotics, sometimes wheels are better.
• 31:47
  SPEAKER_00: Yeah.
• 31:48
  SPEAKER_00: The biggest thing that we've learned from stretch is even within a problem like logistics and even within truck unload, every time you look one level deeper in a problem, there's always an explosion of complexity.
• 32:04
  SPEAKER_00: You might think that, you know, truck unload is a straightforward problem to solve, but it is its own kind of crazy kaleidoscope of complexity that they have to engage with.
• 32:16
  SPEAKER_00: And kind of not appreciating that is uh definitely a way to unfulfilled expectations.
• 32:23
  SPEAKER_00: Thankfully, yeah, we figured that out and we've kind of figured out how to go really deep in truck unload.
• 32:28
  SPEAKER_01: I talked to a lot of physical AI companies, and you know, and there's there's a lot of conversation now around generalized models versus you know uh like platform-specific models, right?
• 32:39
  SPEAKER_01: You know, I mentioned physical AI before, and they're you know, companies like skilled, let's just build this like across embodiments.
• 32:44
  SPEAKER_01: Yeah.
• 32:45
  SPEAKER_01: To a certain extent, maybe Boston IM because it's kind of maybe heading in that direction as far as I see the benefit between sharing a platform in that you can all collaborate together, but beyond that, you know, what is the benefit of having three very distinct embodiments all working on the same platform?
• 33:04
  SPEAKER_00: The benefit of having three different platforms is when you get back to first principles about how those platforms interact with the world, there are things that each one is fundamentally better at.
• 33:14
  SPEAKER_00: Just like if you look in the natural world, you know, quadrupeds versus bipeds, there is clearly a giant niche for quadrupeds in the natural world.
• 33:25
  SPEAKER_00: And that's because fundamentally they have this incredible capability to place their feet where is necessary to maintain balance.
• 33:34
  SPEAKER_00: And a biped with its heavy legs and only two of them will never be as good or as fast as a quadruped in maintaining stability in tough environments.
• 33:45
  SPEAKER_00: And so, like for a data collection platform that spot mostly is, you just can't beat four legs.
• 33:52
  SPEAKER_00: For stretch, yeah, similar thing.
• 33:54
  SPEAKER_00: You're back to kind of fundamental physical principles.
• 33:58
  SPEAKER_00: You're just not going to move boxes faster than a machine built like stretch.
• 34:02
  SPEAKER_00: Kind of back to your question: will those machines increasingly be built off the same learned backbones?
• 34:09
  SPEAKER_00: Definitely.
• 34:10
  SPEAKER_01: What specifically, what is the benefit to that?
• 34:12
  SPEAKER_00: This is a place where like we're just pushing the frontier in research, even.
• 34:16
  SPEAKER_00: Yeah.
• 34:17
  SPEAKER_00: VLMs are kind of lighting the way for the industry, but it's all super, super early for the robotics stuff.
• 34:23
  SPEAKER_00: Yeah.
• 34:23
  SPEAKER_00: The expectation is you should see stronger generalization and stronger reasoning capabilities through the diversity of the data and the training.
• 34:35
  SPEAKER_00: It's just kind of a rule that more diversity is better.
• 34:40
  SPEAKER_01: Like effectively, robots can potentially do more based on the data that they're fed.
• 34:47
  SPEAKER_00: Yes.
• 34:48
  SPEAKER_00: To use your example, yeah, pretty specifically, yeah, we haven't done this research yet, uh, but I would expect that an atlas that has a VLA trained to do sequencing, if you bring VLA data from stretch for truck unload in, you should see a performance improvement from the introduction of that new data diversity.
• 35:12
  SPEAKER_00: Maybe not a huge improvement, but it should be an improvement and not a loss.
• 35:16
  SPEAKER_01: I mean, now this is really obvious as I'm thinking about it, but in effect, having having spot out there, um, you know, doing uh security or uh like inspections in the factory, and then having stretch do a truck unloading, all of those things ultimately they're collecting that data, and that that data will ultimately go to inform and make Atlas a better robot.
• 35:41
  SPEAKER_00: There are a lot of details to get right about okay, exactly what data am I out there collecting?
• 35:48
  SPEAKER_00: How do I get permission to use that data and the relevance of the data?
• 35:53
  SPEAKER_00: Of course.
• 35:54
  SPEAKER_00: Everyone kind of understands this in the industry now, but like being good stewards of your customers' information is actually one of the hardest parts about this whole endeavor.
• 36:05
  SPEAKER_00: We go to really great lengths to make sure that our customers understand what the robots are collecting, how that data is being used, make sure we're really partners in that.
• 36:15
  SPEAKER_00: Yeah.
• 36:15
  SPEAKER_00: But it does make things slower.
• 36:17
  SPEAKER_01: Okay, how about this?
• 36:18
  SPEAKER_01: Uh, here's a place where it makes sense to be owned by uh one of the world's largest car manufacturers.
• 36:24
  SPEAKER_00: Yes.
• 36:24
  SPEAKER_00: Yeah, unequivocally, yes.
• 36:28
  SPEAKER_00: Great.
• 36:28
  SPEAKER_01: Well, uh, Zach, uh, thank you so much for taking the time.
• 36:32
  SPEAKER_01: Uh thank you.
• 36:33
  SPEAKER_01: Great questions.
• 36:35
  SPEAKER_01: All right, that was a great chat.
• 36:36
  SPEAKER_01: Thank you to Zachary and Mia and Carrie at Basa Dynamics for helping set that up.
• 36:42
  SPEAKER_01: Thanks to you for tuning in.
• 36:44
  SPEAKER_01: If you enjoyed the conversation, please support the show by liking, subscribing, and maybe even leaving a rating.
• 36:51
  SPEAKER_01: And don't forget to check out our newsletter over at automated.fm.
• 36:55
  SPEAKER_01: Thanks again, and we will see you soon for another episode of Automated.