The Humanoid Mission in Manufacturing | Boston Dynamics Tech Talk

Welcome. I'm Zach Jacowski and this is Alberto Rodriguez and we're here in the Boston Dynamics studio to talk about the humanoid mission in manufacturing for Atlas and how you build a humanoid brain. I'm Zach. I'm the lead for the humanoids or Atlas at Boston Dynamics. Um I've been at the company for about 15 years. Um prior to leading Atlas, I led spot. You want to say a little bit about yourself? — Yeah. Hi Zag. Um my name is Alberto Rodriguez and I uh I joined Boston Dynamics about four years ago. Uh before that I was faculty at MIT about the time where I got very excited about the prospects of the humanoid journey. So I decided to step down and join the mission at Boston Dynamics. Um I started working focusing on manipulation. Um, and now I direct robot behavior uh in Atlas. Um, where um jointly with my colleague uh Scott Kindersma, we lead the AI strategy and oversee the implementation of that strategy for Atlas. — Cool. I got to ask in the, you know, four years you've been at Boston Dynamics, what has surprised you most about the transition from being a professor at MIT to, you know, going into startup life? — Yeah. Um, maybe one of the things that surprised me the most, uh, and I don't know if it should have, but like, um, I actually have like more day-to-day technical conversations now at BD than I used to have at MIT. Um, not because there's no technical people at MIT, which is many. Um, but I just have more time for it. Like I used to be very busy with teaching and with surveys and just doing random stuff. And, uh, now that's my mission, right? Um, so I I really enjoy that. — That pretty much mirrors my experience coming out of grad school also at MIT and coming over to Boston Dynamics 11 years earlier was when I was in academia, I well I was in academia because I loved building robots. Um, and everything that I had to do that wasn't involved in building robots, I kind of dreaded, uh, going to class, reading papers, writing papers. Um, like I did it all so I could just be wrenching on robots and writing code on robots. And then going over into industry, you're like, wait, I get to do this all day, every day? — That's pretty cool. — Yep. So our vision for humanoids essentially is to build the world's first commercially successful humanoid robot. Essentially solve manual labor in the process. Kind of a big goal. We have a structured way of getting there. You know, strategy-wise, we really believe in general purpose hardware. So we're building a general purpose robot, a humanoid, general purpose hands. And that robot is going to be able to do just about any task in the world. But we're going to start easy. in manufacturing, which is a pretty special environment both because it's a great starting place and it's also uh crazy interesting. We're also going to build a robot brain. You can't really build a humanoid robot hardware generalist without having a humanoid brain that goes along with it. Um, that's going to be the primary topic of conversation today is how exactly we're going to build that brain. — You mentioned a couple times general or generalist. — Yep. — Why general? — So, I've got this video of Hyundai's plant at HMG Metal Plant America up. How about you tell us a little bit about what you see in that plant and then we'll get into why we need a generalist for it. That's one of the many carb manufacturing plants we've visited and HMGMA is one of the newest ones. So, it has the highest ratio of automation. But despite that, once you visit it, you realize that there's still many uh tasks that are still for the most part done manually. And the first thing that comes to mind when you visit that plant is that those tasks are on one hand very hard, very complex. There's tons of variability that comes with adapting with the large variety of objects that go into assembling a car. tons of very dextrous tasks like using tools, assemblies, handling small parts. But if you talk to manufacturing engineers, they will tell you that the reason why they're not automated is not because of how complex they are. They have great solutions, lots of technologies to automate it. It's just it takes too much space. engineering, too much cost to automate each one of those. Uh there's just too many. There's thousands. So special purpose solution for each one of them is just not economically viable. — So there's like the conventional manufacturing automation world where if you want to say bolt a wheel onto a car, you would commission a machine that

bolts a wheel onto that car and that machine probably has a general purpose robot arm in it. But it's still like a thing where you're going to write a specification. and I want my wheelbolter runner machine. And a bunch of manufacturing and automation engineers are going to design that machine for you, make a special end defector for that arm and install it on the line. There's going to be a bolt vibratory bolt feeder and all the all the stuff. — You're going to need tons of uh pro specialized programming to make the machines work. — Yep. Um, and so like when you look at a car plant, we're looking at a bunch of tasks from uh real Hyundai plants here, there's probably tens of thousands of different tasks that need to be done in that plant. And the thing that you really rapidly get to is if you're going to design automation for each of those things, uh you'd be uh spending way more than it's worth to automate those tasks and you'd probably be well into the next century. uh like an average estimate of like how long it takes for any one of those um integration efforts um it's in the order of about a year and north of a million dollars for any step that you want to automate. That's sort of the state-of-the-art today. — Let's talk about how the tasks that we see in a normal car plant breakdown. um you kind of have well really simply I would say you have material handling tasks and then you have assembly tasks right in those material handling tasks like the plant words they would call those things like sequencing kitting racking moving material from you know the warehouse to line side and then assembly is like we were talking about attach the wheel to the car and put the bolts in and all that stuff. What are the high points of those tasks to you? — I think that I mean while those are clearly different the level of dexterity that is necessary to execute them is different. One is mostly about picking and placing and inserting parts in containers while the other one has to do more with very dextrous part handling. The biggest source of complexity is the large variability which is shared by both right like you were saying a car has like tens of thousands of parts uh all of them different and in the same assembly line you might assemble between five and 10 different uh types of cars each one with its own trim and each one with like 10 to 20 different colors — and you turn over the model year every year so you change a bunch of trim parts interior technology and stuff and then you do a major model line refresh say every 5 years. — So it's that variability that really um puts a ceiling to um like traditional automation right how what the core question is how do we make something that is more flexible that provides uh the degree of generality so that you can automate those tasks. — Yeah. So how do we make something more general? — That's a trillion dollar question. Let me maybe backtrack one second before answering how we make something more general uh and describe that is more specialized. It's not that long ago. If you backtrack like four to five years, this is how robots even humanoids were programmed. — Y — right. So you basically have three to four core workflows that are necessary. Like you model your assets, you model the parts that you have to manipulate, you model the environment that you're going to operate in. You build uh perception systems that are capable of recognizing and perceiving those parts and those environments. Then you build skills that are capable of handling those parts, manipulation skills. Like let's say in a logistics application it would be picking those parts or extracting or inserting those parts from the containers where they're contained and then you build on top of that an agent that is capable of going sequentially between the steps of the task like — so yeah and when you say build the behavior that you made it sound pretty easy but like what is the actual work in doing something like you getting a robot to say you know, pick we showed a video of moving engine covers around like what is involved in getting a robot to like reach in there and grab that engine cover. — Yeah. In practice, what that means is you have engineers that have the right intuition to program how the robot should approach those parts and um what's the right way to let's say position your body with respect to the environment to be able to reach far enough to extract it to hold the weight. So there's a lot that there's a there's

a cycle that has to do with uh using um human intuition in the form of uh [snorts] expert robot programming. — Yeah. — Testing uh finding failures, iterating and uh getting that system to u have high performance. — Cool. — Now there's obviously like uh an issue with that model which is a scalability, right? Like that's a system that if you have to deploy three times, you can do it. If you have to deploy thousands of times, uh doesn't scale fast enough. — And it's a system that also doesn't improve over time, doesn't accumulate experience, uh knowledge, uh to make the next time easier. — So that obviously isn't going to work. Uh so we've built a robot hardware generalist and the world you described there has a large team of highly credentialed robot engineers program it to do each task just you know not going to work at its face. So we've decided we want to build this uh robot generalist or we'll call it a robot brain y — uh because that's essentially what it is. How do you build a robot brain? I think that the core part of it, it's still pretty similar to that cycle of refinement. — The core difference is that you don't want to program the robot explicitly. Instead, you want to teach the robot how to do a certain task. And as the robot is doing it, you want to tell it when it has failed, how to correct for mistakes, how to improve its performance over time. So there's a core workflow there. Um something that we refer to as post training. Mhm. [clears throat] — where basically you demonstrate the robot how to do something and um a certain number of times could be it could take in the order of a day for example of demonstration of an expert demonstrator uh to tell the robot how to do that task and how to um correct for certain errors that might happen during the execution and accumulate all that knowledge. now not in the form of a program an algorithm but rather in the form of uh knowledge ingrained in a neural network. — Okay. So that post-training step that you just described there um you said you know uh some number of demonstrations like let let's say generously takes a day to learn how to you know grab that engine cover and move it to the other slot. That's still a awful lot of days. uh to do these you know 10 tens of thousands of different tasks. So like what are we going to do about that? — So um the most um interesting part of this new strategy is what we call pre-training which basically accounts for a um a system that makes an good initial guess as to what should be that policy. So that once you go to factory and do the first deployment, the behavior is already somewhat good and it takes a small amount of time to improve it. Now the big question is how do you build that system that is capable of producing a good initial guess that has accumulated enough common sense and general understanding of what it means to move and manipulate objects and do tasks like manufacturing tasks so that is capable of that. And so what you're describing there isn't like some unsupported fantasy like that. That's essentially how modern LLMs work. So like this is how something like chat GPT is built. — Yeah. Exactly the same way. Uh there's core differences because of the complexities around actually generating and finding the right data that is capable of providing understanding of visual spatial relationships, dextrous manipulation, agile whole body coordination and motion which is not something you find in text or in audio or even in video that is specialized for robots but the techniques the technology is essentially the same. So something like chat GPT is built on kind of the sum of all human knowledge on the internet. Y — um we don't have the sum of all human uh physical behavior available somewhere. Uh so how do you get that? — So um there's two parts to it. Um the first one is um we have um what I call data swim lanes that generate good data that is useful to capture that general understanding. Teleoperation is one of them. So scaling up tele operation so that you don't solve one task but you solve hundreds of them thousands of them not necessarily during deployment in factories which is more costly and more difficult but in a dedicated space where you do that. That's one bet and to some degree that's what most companies are doing today. — But then there's other very interesting

directions which I would qualify closer to research efforts today that provide ways to scale up data generation much faster. One of them is reinforcement learning in simulation uh where you don't learn necessarily by providing optimal demonstrations but you let an agent to explore by itself trial and error and optimizing behaviors since we can scale simulation to a much larger degree that can provide a way to increase data generation. And the third one is directly by observing human behavior. If you design a robot that somewhat mimics uh or somewhat gets close to how a human operates — Yeah. — in its form factor uh in its hands, you can um learn um some degree of common sense of physical uh behavior from observing humans. And then there's a second way by which um we can accumulate uh knowledge at scale in this pre-training phase which is what we normally refer as the fly the data flywheel right so once you've done this for enough tasks and you get to deployments uh the data that gets generated by those deployments is very rich is uh data that actually represents the environments and the tasks where you want your robots to operate. So that generates um information data experience that goes back into pre-training um in a way why it's important to get in the short term to be able to do deployments even if it's within without robots being yet or the robot brain yet being perfect. — So that's the process that we want to execute and those are the kind of data inputs we need. Mhm. — There's the matter of like what kinds of ML models are we training here? Um there's like a lot of different things floating around you. There's LLM, there's VAS, there's VLMs. Uh do you want to walk us through a little bit of a map of kind of how robotics was previously in the world we're headed to? — Yeah. Um so there's like a spectrum, right? So classical robotics like four years ago. We build specialized models for as I describing earlier for perception, specialized models for high level coordination, specialized models for manipulation, — even specialized models for mobility. Now we've agreed that maintaining and uh evolving all of those models with all that complexity is not the scalable process. On the other end, you have a fully end to- end solution where robots ingest pixels uh or raw information from sensors and at the output you produce torqus or currents that go to the actuators. — Yeah, — we know how to train uh these more abstract systems through experience. The complexity is the scale of data that is necessary, right? And like LLMs today are a perfect example. If you have enough data, um you can rely on uh raw representations like not just LLMs but like video model prediction — uh systems, right? You can rely on raw representations to train really highly performance systems. — What's the degree in between where an actual practical solution is going to land on? Uh for example, for manufacturing, um it's still a question, right? like you might want to benefit from the fact that manufacturing provides a little bit of structuring so you know which objects are going to be ahead of time. You know um the uh furniture and fixtures you're going to encounter with. So you might want to put that into pre-training. — Yep. So in that the spectrum from fully modelbased to fully end to end trained uh there's different progressions of abstracting more and more layers of like perception um high level cognition um dexterity uh and the mobility. — Got it. So like in the large behavior model video and the associated paper with it like where in that progression is Atlas at that point in time. So that's a great example. That is not a system that ingests pixels and outputs torqus, right? It's it outputs commands to more abstract representation of what a robot should do. So it provides commands for what the endector should do, what the feet torso, what the hands should do. Yes. — Hands, feet and torso should do. Below that we have a very powerful whole body controller that consumes that and realizes that on the robot.

— Y — right. So um that provides a layer of abstraction that makes learning simpler and reduces the amount of data that we need to generate but still allows us to exploit the high agility and strength of a platform of a humanoid platform like Atlas. I mean, that's tricky because like that that's not just a technology progression thing. Like there there's real frequency separation between what Atlas is doing like with its actuators versus like looking at a scene and deciding what you need to do next? Like is it necessarily a given that should be one neural network? — No. Uh absolutely it's not a given that it should be just one single neural network. In fact, it is pretty common for especially dynamic balancing robots like humanoids or quadrupets to rely on what we call system one, system two separation where you have a system at the lowest level which takes care of whole body control. That the system that needs to run very fast and needs to know and understand sort of the extent of the strength and balance of the robot. It needs to know that you shouldn't move your hand too fast over in this direction without counterbalancing. — Sure. — So that's like your in humans that'd be like your cerebellum in your nervous system. — So it's unclear that you might want to force everything else in your system that is in charge of acquiring common sense to also have to learn that at the same time. — Yeah. — So there's some value from separating that. — Let's jump in and you know talk about some of these data sources in a little bit more depth. Mhm. — Um so let's talk about teley operation. — Yeah, I would say that teleoperation is today um the the most um valuable source of data for early deployments or for early behaviors, right? So if you want your robot to um uh exert a certain behavior in the next couple weeks, teleoperation is your friend. Um it's the most uh valuable kind of data because it's fully embodied, right? So you teleoperation means that you are controlling the robot and the robot is experiencing what it means to do a certain behavior. So the data that you collect that experience is through the body of the robot. — Um so in a way there's like zero gap between the demonstration and the actual data you're collecting. — Why is that important? That's useful because um that kind of data is not polluted by artifacts of the inter your interface to the robot. — Okay. — Um it's like ground truth data, right? Like if it worked, it means that the robot will be able to do it again by using the same commands, sending the same signals to the actuators. — So that data is valuable. um it's probably the hardest to scale because it requires um building a very proficient teleoration system right so our demonstrators at BD um they learn over a few weeks they become experienced in understanding the extents of what Atlas is capable of doing by spending hours um moving robots through this telebration system which basically is in the form of a BR system where they see through the eyes of the robot and um they have trackers in their body so that the robot reacts life in real time to the same motions that they're doing. — Mhm. — Um once they become experienced in doing that, they are free so to speak to command the robot and get it to do anything that teleoperation interface allows them to do with its observability. — There's a the scale thing going on here is important. So like we're talking about trying to build something like an LLM for physical movement of a robot. Um again those are trained on close to the entirety of human knowledge. Um and we're talking about collecting you know thousands tens of thousands maybe you know generously hundreds of thousands of teleyoperation uh you know motions and hours of data. Mhm. — There's a immense scale gap between what we want out of these models and that kind of data source, right? — Yeah, that's correct. And um it's not the secret, right? That uh that's probably not going to be sufficient. Um, we think that it's still very valuable because of that sort of zero gap transfer between the data we're collecting and the behavior, the motions that the robot can execute — has tremendous value when it comes to demonstrate um embodied behavior. So if you want the robot to crouch, if you

want the robot to reach, if you do whole body coordinated motion, it's a great way to generate high quality data. — Yeah, it's definitely going to how we're going to pilot the giant Atlas to fight Godzilla. — Yes, exactly. — Um, will you be inside? — No, no. Someone much better at it than me will be. All right, let's talk about that second kind of data then. So you mentioned reinforcement learning or trial and error for manipulation. What's going on there? How does it work? — So reinforcement learning in the scale of high quality data um it would be below teleoperation because now you have this sim to real domain gap. The data you generate in simulation um is doesn't translate perfectly to the robot. um we're pretty good at Boston dynamics of understanding um that seem to real gap but there is still a gap — however simulation gives you the option to scale massively the data generation workflow um which is a great benefit but also allows you to control the scenarios where you're learning right you can expose the robot to small variations of anything you want in a control manner oh you want to grasp this object well do the same thing but also if the object is one centimeter to the right or one to the left or in thousands of variations. — Mh. And um in that context and because of the massive scale, we have access to algorithms that are very hard to replicate with real ground and real data or with real hardware, which are algorithms that are based on trial and error and optimization to fine-tune behaviors that become better and better over time. — Um that can become even better than demonstrations. um just because they're trying to optimize maximize some objective. — Yeah. So I guess that that's an important thing that you we didn't quite hit on earlier is with teley operation you always have this interface that makes it a little hard to do really precise, really fast, really dextrous things. Um you're saying that that's really not a thing for reinforcement learning. Yeah. Um actually that's a pet peeve of me uh that uh I have with teleoperation which is that we use teleoperation to provide high quality demonstrations that are meant to build these physical understanding of what it means to move in the world. However, um demonstrators unless they're very skilled um because of the complexities of the teleoperation interface, they end up demonstrating behaviors that are somewhat subpar. They are like kind of slow or too sequential. Instead of using dynamic behaviors, they use quasatic slower behaviors. And it requires conscious design of the teleoperation interface but also conscious training of demonstrators to avoid that. — Mhm. — Simulation or reinforcement learning doesn't have that limitation. You can sort of optimize the performance of um a any given task to its maximum. Um you can force it to be fast and dynamic and robust by exposing it to u many variations. And we've had like um we we're very happy with the success we're having in that direction in particular for Dexus manipulation. — Yeah. So we have some pretty cool examples here of three tasks that you're uh that our reinforcement learning folks just trained up recently. You want to talk about those three? — Yeah. So you have here examples that are sort of like the bread and butter of what it means to do work in the assembly line. So, insertion tasks that are haptically driven like uh inserting um a steering wheel in a socket or inserting a plug into small part into a jig. Any of those tasks um is very hard to teleoperate either because it requires small very controlled finger motions or because it requires sensing haptic signals that are not directly visible. And so you like looking at these look like specialist policies like you trained a thing that specifically only knows how to pick up a steering wheel and put it on the steering column. How do these end up part of this robot brain we're trying to train? — Yeah, that's a fun question. So that's correct. Each one of these is its own individual policy. We call this a specialists. They're really good at doing one thing with its domain of generality or robustness. But is that one thing and we're very excited about

the fact that we can scale that process pretty seamlessly of creating thousands, tens of thousands, hundreds of thousands of specialists. In a way, these become your teleoperated examples, right? If you do rollouts of these policies, this generates a kind of data that then you can uh condense into a generalist like you would do with behavior cloning uh teleoperated data. — Cool. And then it not just manipulation but also the locomotion end of Atlas too. — Yeah. Absolutely. Um right. So locomotion especially agile locomotion if you want to go fast do natural walking behaviors cartwheels suffer from the same issue. That's very unrealistic to expect we're going to get one of our demonstrators to get Atlas to do a cartwheel. — Yeah. That's something that you really want to optimize through because it's a dynamic behavior that has very contrived regions of attraction of where you need to get the robot. You need the robot to accelerate to a certain velocity so that it can jump and do a cartwheel. It's not something you want to do other than in simulation. — And then this third kind of data, human demonstration or you know human observation data, what does that mean? — That's the biggest bet I would say. um is the more uh long horizon bet um but also the one that uh has highest potential of scope. So um it means anything from like all the way to learning directly from YouTube videos, right? So, can you get robots to learn what it means to interact with the physical world um with dexterity and common sense by looking at videos of people repairing bicycles in their backyard? Um or looking at videos of people constructing things. Um — so that's the extreme of it. But there's a spectrum of ways in which you can use more um open-ended demonstrations from people that uh doesn't have to be all the way to teleportation. For example, something that we're very excited about is echocentric demonstrations where you get people to wear sensors that make it easier to capture what the person is doing. It could be the same sensors in the head that Atlas has in the head. Um, it could be gloves uh on their hands that capture tactile signals and just get them to do by their own life or their own job and just learn um what it means to do that job uh from that data. — Cool. And that does sound like or rather the reinforcement learning and that one do sound like the kinds of things that conceivably could get to these um scales of data that are being used to train LLMs. — Yes. Um agree both of them and um I would say that it's not necessarily either or each one of those two has their own strengths — and weaknesses right so clearly simulation um from our experience is a really good choice for u dextrous manipulation — but it might be more difficult to learn things that involve um highly deformable interacting with highly deformable objects for example uh because they're just naturally difficult to simulate or require a lot of compute. Direct echocentric demonstrations with people wearing sensors. Um they're great at capturing just like the breadth of common sense of what it means to interact with objects and the sequence of things that you should do to accomplish a certain task. But it will be more difficult to learn directly from that dexterous manipulation because they're not embodied in actual the body of the real robot with its own sensors. — Cool. Okay. So we got a complete plan to build a robot brain. — Yeah, we should. Now we just need to execute. — Um so let's Yeah. So like let's go back and recap like so what does it mean to build a generalist and like what are the essential uh characteristics of that thing uh or what are the characteristics that thing will have to have for us to be successful in putting humanoids to work. I would say that there's like four core tenants, right? So, general purpose hardware and you've talked about it. Um, ease of retasking, very important. Um, to avoid the issues with the cost of integration of every sing uh every single task, ease of retasking and getting to a point where it's not months, but rather it's days or hours what it takes to get a robot to do a

certain job. Yeah, — natural interfaces for interacting with a robot rather than requiring expert programming. Um, as and our customers will want a more natural way to tell robots what to do either with natural language or with direct demonstrations. And the fourth one, safety around people. — Did you get reliability in there? — Uh, yes. Um, I did not get it. But — so we got five. Let's Okay, let's say five. — Yeah. So, we're building this general purpose robot. robot brain. In conjunction with that, one of the really cool things about working with the Hyundai group is we're not just building this robot in a vacuum. like we're also redesigning car plants and redesigning general manufacturing facilities alongside understanding what Atlas's capabilities will be. Can you tell me a little bit about what's exciting about that to you? — Let's see. One thing that um Okay, maybe two things. — The first one is that I think working with Hyundai has made the mission very clear. — Yeah. — Right. like our mission is to transform manufacturing and um once you know that that's clear it's um it relieves a lot of pressure and stress of like actually thinking and dreaming about any other application domain where you could also be working which there are many and which we might go also after but what's non-negotiable is that we need to solve and transform manufacturing so that has been uh Um I think exciting to me the fact that we have a clear mission which is very hard but it is clear. — Yeah [clears throat] that I mean that was honestly one of the hardest parts of the early days of spot was yeah you had this amazing robot and we had so many application ideas for you know what it could be really useful for. Like today we look at spot and we see it being used all over in these industrial inspection applications and it seems obvious now but like in the early days we were spread really thin between so many different applications testing and trying to do demos of each of them. It's really refreshing with Atlas knowing that you know if we can just solve this manufacturing problem that we can be successful and then we can go into the rest of the stuff. You know we can we're not going to stop at manufacturing. We're going to do curbto delivery. We're going to put robots in your home, clean your tables, make your bed. But we know which one we have to do first. — Exactly. Yes. Yeah. Yeah, I mean um at the much smaller sca smaller scale personally also um like before joining BD I was faculty at MIT and as faculty um the journey is always about finding the right motivation to solve the problem that you think you should be solving right like you always have to convince yourself actually am I solving the right problem or am I confused should I should be solving a different one — um so the change to BD has been refreshing also because of that like I don't have to ask that question myself anymore. I know what problem I need to solve. — Yeah, I feel the same way. So, part of the reason why we're talking about this today is we need more folks to join us on this mission building a humanoid robot brain. Can you tell me a little bit about what it's like day-to-day on our machine learning team? — Yeah, absolutely. Um, like that's a larger area in Atlas where we're hiring people with experience on building large behavior models. um and people with experience on reinforcement learning. The experience in the team I would say is um it's hybrid right so there's some people in the team that are very uh excited about research and they spend their time you know uh solving the problems that uh we know we will have three years from now uh or five years from now. Um but there's also people in the team that are very uh enthusiastic about uh how can we get as fast as possible to deploy these robots for real and that um also people that need to know need to be very experienced on uh these modern techniques like behavior cloning like efficient um data collection techniques like uh efficient co-raining techniques uh to fuse data between simulation and tail operation — that also are good robotic ists, right? Um, so the experience I would say that it's um on both sides and we're always looking to hire good people. — What do you think your team looks forward to and enjoys the most

dayto-day? — I think the team is very excited about um I don't know if the team but myself I am very excited about the prospects of really enabling Dexter's manipulation. Mhm. — I think that is the ultimate frontier of what it means to do general purpose work. Um and um BD is a place where you get to experience both the design and fabrication of um high-end um Dextrous grippers and you get to experience the development of the techniques that are going to get those grippers to do very cool things like using tools like handling objects. Um, and to me experiencing that and seeing that sort of rapidly progressing is fascinating. — The uh I got to answer this myself. I have so much fun. The you know the really special thing about robotics as opposed to any other place that you could be working in machine learning or really in any of the functional disciplines is that like we do the whole thing. Like so you mentioned, you know, folks at Boston Dynamics design grippers and like if you're a machine learning researcher, you get to talk to the folks that design the fingertips dayto day, argue with them, you get to see your stuff show up in the hands 3 months later. Um, and it extends all the way through the other side too. So like we have folks who work on robots daytoday, diagnose things. We also have folks who, you know, sit 15 ft further away from the robots in the lab who think about how do I wrangle a thousand GPUs to run this training job faster. Like you see every aspect of putting a robot together all in one room. — Yep. — I don't think it's there's very few places in the world where you can get that. — It's very exciting. — Well, thanks Alberto. Thanks for telling us what's going on the ML side of Atlas. Um, and we'll uh I know we'll have to see how fast uh we get to that full generalist. — Yes.