Sim2Real and Optimus, the Humanoid Robot with Ken Goldberg – 599

All right everyone welcome to another episode of the TWiML AI podcast. I am your host Sam Charrington and today I’m joined by Ken Goldberg, the professor of industrial engineering and operations research and the William S Floyd Jr. Distinguished Chair in Engineering at UC Berkeley. Ken is also the chief scientist at MB Robotics. Before we get going be sure to take a moment to hit that subscribe button wherever you’re listening to today’s show. Ken it has been a bit welcome back to the podcast. Thank you Sam. It’s a pleasure. I’ve been enjoying listening to your shows over the last couple of years. Awesome yeah it is hard to believe that it has been two and a half years since the last time we spoke. Of course a ton has been happening in robotics and I guess we’re going to use this next next little bit for you to catch us up on everything. Well it’s been actually an amazing time. I mean with it with a pandemic obviously you know we were in the middle of it when we last talked and so much has has happened. I would say that it’s actually been a very productive time for robotics that people have made enormous amounts of progress. There’s a lot of publications, a lot of research that’s been going on and also the world has changed in interesting ways that I think is as is favorable to robotics as a field. Absolutely absolutely. Now I’ll refer folks back to our conversation that was episode number 359 back in March of 2020 the third wave of robotic learning for a great conversation and your full background but for those who you know haven’t caught that why don’t you share a little bit about how you came into robotics. Oh well I was I’ve been interested in robots since I was a kid and back in the old days of Star Trek and and was in space and I’ve just continued that for you know 50 years and what I have to say is I rediscovered robotics as an undergrad and found that there was just absolutely a fascinating set of questions. I was lucky to have a great advisor, Ruzuna Bicci was at Yupen at the time she took me under a wing really mentored me and she’s still a good friend. She was at Berkeley for many years now back at Berk at now back at Penn and then I went to Carnegie Mellon where I worked with both Matt Mason and also for a little bit with Mark Rayburt and so I had just the opportunity to work right at the heart of where robotics has been had been start growing and I took an interest in one particular problem which was grasping and I I’ve been working on the same problem for 35 years. Which says a lot about how hard that problem is. Exactly exactly you know it’s interesting because most you know I think everyone on this who’s listening to this knows that that it’s a it’s a hard problem for robots but it’s still interesting for the public that most people you know we humans do this effortlessly it’s very easy to pick up almost anything right that we you’re handed in fact my dog can pick up almost anything with a parallel jog ripper right it’s it’s jaws and it’s quick as well you know you can pull out some very weird shaped object and it’ll quickly figure out where to grasp it that isn’t fascinating skill but if you do that in front of a robot it is we’re very far from being able to do that reliably. Yeah I think one of the things in kind of looking over some of your recent work that most jumped out at me as indicative of the kind of progress we’ve made is a paper that you worked on with your team autonomously untangling long cables and I guess that struck me because I remembered back to I don’t I must have been maybe four or five years ago I don’t think it was it was your paper I think it was Peter Abiel had this paper of like trying to untie a knot on like a gigantic rope you know a single knot and if the color of the background change didn’t work if the color of the rope changed it didn’t work and here you are with this paper like untangling a hornet’s nest of cables. Well it’s not quite the hornet’s nest yet but but you’re you’re you’re you’re right in that that’s that’s a very good perspective we were Peter was working with a thick cable and it was I believe that was tiny knot okay to get to create a knot a few years ago we started looking at untangling and we started with very simple very small sections of cable so on the order of eight inches and we actually used our surgical robot as an implementation of that so we had a da Vinci trying to untangle these very small segments and they were very simple knots they were just overhand knots and that that was a that’s a hard problem because it’s you’re in the realm of deformable objects and well you know and there’s an infinite state space for those there’s also self-occlusion and and and in fact though we minimize the self-occlusion because of the the size of the knot the size of the cable right it was fairly small so you didn’t have all this overlap and slack if you will so the key was to identify where the knots were which we we learned with a deep network lots of examples and then it would just start pulling at these knots and then and then be able to open them what was what was exciting was the the we started thinking about longer cables and really extending this into the macro scale with a full scale robot a dual armed um um um you me robot and really starting to think about all the complexities now of of of really trying to untangle this this thing, pull it apart and manage the workspace for a dual-armed robot, which has got a lot of complexities in its own right, because you have to avoid self-collisions with the two arms. They have to work around each other, and there’s challenges and perception. And also, the workspace of the robot is surprisingly small. It’s usually about the size of a dinner plate, really. And the cable, in this case, was about three meters. So that greatly exceeds the size of the workspace. So you have to think about how to move things in and out of the workspace and resolve invigilities. But that has been a very fun project, I have to say. The team has done a great job. It was actually led by two undergrad who became master students now, Vynavy and Koshek. And with Justin Kerr, a PhD student, they presented this at RSS this year. And they did such a fantastic job. They won the Best Systems Paper Award there. Yeah, congrats to you and them on the Best Systems Paper Award. The award calls out the systems and nature of this. Can you talk a little bit about this as a systems challenge? Definitely, and thanks for asking about that. In fact, it is a systems paper, because you have the system of the perception system, the planning system, the actuation system, one of the key things that made this new work possible was a very, very clever hardware design that Justin came up with, which was to add a little foot onto the parallel jaw grippers. And if you think of it as just a little L-shaped, but a very small, I would almost a toe, if you will. What that allows it to do is that the grippers now can be fully closed, and that holds the cable tightly. But if you open them by a little more than the diameter of the cable, then those two toes basically are interlock, and they prevent the cable from escaping. So we call that caging in robotics. Caging is where you have the object contained, can escape, but it may not be held immobilized. It might bounce around. So caging, like you put your hand or bird in a bird cage, it can move, but it can escape. So caging is an interesting geometric problem in its own right. But here we think about caging, which is a way of allowing the gripper to enclose the cable and slide along the cable, which tends to pull through knots and tangles. So that turned out to be a very, very critical part of the system, that piece of hardware, that allowed us to introduce new primitives. So what we call cage pinch dilation, which is where we pinch with one jaw, cage with the other jaw, and then pull dilated. And that allows the system to untangle individual knots. And then we systematically pull through the slack. And there’s one other challenge is the depth finding where to grasp a cable is difficult, because the cable is not lying perfectly on the plane. So it tends to loop. And sometimes those loops can lift three or four inches off the plane. So if you go to grasp them, you really do need to have some sense of depth. And so we use a depth camera, but those are very slow, because of the scanning process. So what we are moving to right now, and this is continuing, is trying to avoid the depth camera. Is the depth camera that you’ve been using is it a vision-based 3D stereo thing or more like a point cloud connect type of camera? More like a point cloud connect. So it’s using a laser scanner. And it’s a structure of light. But that scan is slow. It’s about one frame per second at best. And it also has a lot of specularities, problems with it, just getting that accurate is challenging. So here’s what we’ve been thinking about, Sam, is trying to imagine that we can’t know exactly the height of the cable. But what we can do is we can use geometry and physics. So what we do is try and estimate a position for the gripper above the cable in such a way that the cable is somewhere in here. We don’t know where. But what we’re doing is now lowering the cable down in such a way that it will again cage the cable as it moves to the work surface. And then it closes. And so we’re guaranteed to get the cable, even though we don’t know it’s exact depth. So this is the kind of thing we’re interested in, is how to use geometry and mechanics to perform things where the centering may not be sufficient. And another example is that if we, here’s something we’re very interested in right now, which is starting to use the subtle clues, where I want to determine where, let’s say, I do what really care about where the cable is. It’s sort of moved up and sitting on the over the table. What I can do is move down with the jaws and carefully monitor what’s going on. And when the moment when there’s a movement of the cable will be an indication that I’ve made contact. So the vision can actually provide a lot more information than we tend to think. We don’t need a contact sensor here. We have a vision because the cable will imperceptibly move just a few pixels. But knowing that is a trigger and say, wait, that’s the height. And we know the height of the robot arm because of the kinematics. So I can determine that, well, therefore, that’s where the cable is. Interesting. Interesting. So hardware played a big role in allowing you to do this. Can you talk a little bit about the software or MLAI side of things? And with a particular emphasis on the things that have changed over the past couple of years that have really helped to tackle a problem like this? Sure. And because we’ve been, this problem has been evolving. And it’s one thing I actually just want to mention. I really believe in. So I came to the students last week and were saying, I’m really proud of the projects that tend to continue over multiple years. DexNet, as you know, was our grasping work. And it started with DexNet 1.0. And when all the way up to 4.0, and we’re now working on 5.0, actually. Because what I like is that we take a problem and we really try to dig deeper and deeper into it. And in this case, it’s really looking at the failure mode. It’s really trying to not be satisfied with the performance that and wanting to understand, how can this be made better? How can we really push the envelope? How can we really reduce the failures? In this line of work, we systematically characterize the failures. So just as an aside, I see today so many papers that say, we’ve big the state of the art. And thank you. This is a superior and end of story. And it’s very frustrating because I always want to know, well, where, how did you beat the state of the art? Just giving me a mean success rate doesn’t tell me a lot. I want to know, did you succeed on a number of cases where the baseline’s failed? Or what did you succeed? And where did you fail? And that failure, it’s interesting, because I think an instinctive desire not to want to look at that. But it’s actually where the most interesting aspects of the problem are. And so you really want to study those failure modes. And so we do. We characterize them into different categories. Every single case, we really determine what was the failure mode. Where did that one go wrong? And then we try and look at those and say, OK, how can we address that? So the idea is that in each case, we had to really develop new primitives and new primitives both for perception and for action. And then we try and think about how do we sequence between those primitives? So these are the software aspects. One of the algorithmic aspects is learning to recognize just where knots are in images. And so we did that in real by sampling lots and lots of examples. We just have the system watching the cable. And we do a certain amount of self-supervised movement. So the cable, basically, tie a knob in it. And then we allow the system to move itself, move the cable around, taking images over and over again. And then we either manually or what we hope to do is to, more and more, to have a self-supervised method that provides ground truth, labels images, and then we can train a network. So that’s been a core part of it. And that’s a huge thing where deep learning has changed the equation over the last decade that we have that perceptual ability that really can solve some of the very, very complex perceptual problems, where it’s very hard to analytically determine what you’re looking for. And this is to determine what is a knot. It’s actually very hard. One thing that also is interesting is there’s a whole body of theory called knot theory. And mathematicians have been working on this for centuries. It’s very interesting. Where they begin by turning it into a graph. So they abstract away from all the geometry and the physics. But I’m very interested in combining these. And there’s one technique that comes from knot theory called rate of mice to move that is actually analogous to what we do when we pull through the cable from end to end. So we’ve been applying that. We’ve been applying new forms of learning in particular. We’re right now, again, trying to remove the reliance on the depth sensor and learn permittives that can determine, that can compensate for not having depth. And I’m also very interested in having human in the loop. And by that, I mean very sporadically when the system is stuck, where it’s not making progress, or it determines that it’s in a sufficiently uncertain state, it can call in a human for help. And this is interesting in its own right, in more general sense of when, how do you do this? You want to minimize the burden on the supervisor, the human. And by the way, this is a widespread issue, for example, as you know, Google is testing an automated car service, taxi service in the Bay Area. And my understanding is that they have humans networked in and are standing by. Now, you have one human, probably controlling multiple taxis. So now you have a question. How, when do you call that human in? And you want to minimize that, because ideally, you can have one human supervising 50 taxis. So you don’t need the person that often. But in any robotics case, it can be very tedious to be constantly bothered. So there’s a nice problem in when do you call human? And also, when the human comes in, when do you transfer control back to the robot? So you have these nice dual problems, and they both have to do it in a sense, a model of confidence. So we’ve been looking at that. And I think that that applies to many of the tasks we’re interested in, where there are these failure modes that you’re really unrecoverable, at least currently. And so that’s where there’s no harm. And maybe once an hour or so, you want to have a human come over just to just something, and then go back to whatever they were do it. Yeah, when I introduce you, I referenced that you’re in the School of IE and OR, and it strikes me in your description of this problem. There’s also some interesting kind of classical OR queuing theory types of problems in there. Despite the degree to which I enjoyed working on that kind of stuff in grad school, I have not looked a whole lot into what’s happening to marry machine learning and queuing theory. Do you know of anything interesting out there? Oh, well, there’s quite a bit. I mean, queuing is also a beautiful model. There’s always use of Poisson distribution assumption. And a lot of nice theorems you can prove on that. But the reality is not, it doesn’t behave that way. So how can you generalize that to real empirical or real empirical distributions? And I’m glad you mentioned that you know this connection with queuing. And in OR, my colleagues say, well, we’ve been doing AI with the century now. I mean, because in some sense, mark up decision problems, these have been the core of operations research for a very long time. And so those are early forms of AI and still and being rediscovered in a way, especially where it comes to optimization, which is at the core of deep learning. And so many of the models that we’re using now. So it’s very natural to have connection between OR and AI. And the industrial engineering side of it comes with the other aspect, which is, how do you make these systems practical? And that’s where factors like what we’re just talking about, the human interface come into play. There’s a distinction between robotics and automation. And robotics is obviously much more popular and enticing and the press loves robots, et cetera. And so I’ve always been amused by the fact that if you start talking about automation, it tends to be just that sounds like something in a factory. I don’t want to really talk about that. But if it’s robotics, it’s really exciting and energizing and it feels like science fiction. What’s been happening, I think, in the last few years, is that there’s a trend toward automation because there’s a recognition that we want to start putting things into practice. And that’s where you have to worry about robustness. You have to put guarantees on performance. You want to worry about cost, reliability, all those factors that are often overlooked when you’re just doing something in a lab. Yeah, interesting. I thought you were going to go a totally different direction with that last comment. People often will talk about software robots. And I’ll ask you what your take is on that. But to me, a robot, part of what fundamentally defines a robot is this bridging of the digital in the physical realms and something that purely exists in the digital realm. Unless we’re talking about a simulation of something that exists in the physical realm, I don’t like calling those things robots. The software or it’s automation supposed to software robot. Oh, no, I couldn’t agree more. I mean, I think that’s actually, it’s a misdomer. People often say bots. Oh, it was the bots took down this website because it was automated, some automated modules that would be able to do something, but they’re just software. And I think that is definitely a confusion. And to my mind, the robot has to have a physical component. In fact, this brings up another aspect, which is a lot of research has been done just in simulators and then demonstrated with simulation. And I think there’s a danger there that if you can almost have a self-fulfilling prophecy, you build the simulator, you work with a simulator, you tune off of the simulator. It’s very nice because it gives you the ability to collect lots of huge amounts of data, and you can do resets in the simulation. But if your simulation is even slightly deviates from reality, when you now take that policy and put it into practice, you have a performance can degrade dramatically. And so a lot of the early New Joco demonstrations of walking machines, et cetera, looked great, and they look beautiful, and I’m just surprising how fast they would learn. But then they would not easily transfer into real machines. So this is the sim to real gap that I think is so interesting right now, and is really important to recognize. And you’ve been doing a bunch of work in that area as well, you and your live feed talk a little bit more about how you characterize that sim to real gap as a set of research problems and some of the specifics that you’ve been working on. Sure. One of the things that I’ve always been interested in is the limitations of simulation in grasping. And I talk about the very real problem of being able to, there’s essentially indeterminacies in physics that are due to friction. And the example I always like to point out is just pushing a pencil across your desk. And if you do that with just put your index finger and you do that repeatedly, the position of the pencil will be very different. And so it’s a chaotic system. It’s basically based on the complex surface physics, the surface topography. And that is very difficult. It changes every single time you perform this. So in a sense, it’s impossible to predict how that pencil is going to, with the final state of the pencil. It’s an undecidable. I feel like I think we talked about this in a further amount of detail last time. All right, good. You have a good memory. I know I don’t want to repeat myself. But it’s good to know that I wasn’t saying that because you’re repeating yourself. I was more saying that because there’s a part of me that wants to get into a philosophical argument about it. And I’m wondering if I got us into that philosophical argument a lot at last time. The basic question being, is it practically chaotic because we don’t have the resolution to incorporate the fluctuations in the surface and the dust particles and all these things? Or is it, if we could capture the microscopic physics, would we then have a deterministic system? Or were there always be some element that we can capture, humidity, temperature, what have you? No, I love that. If we could probably talk for an hour just on that, I’ve been using the term, thinking about the terms epistemic and aliatrary uncertainty. And this is exactly what you’re talking about. So the epistemic is that we just need to model this better. We have those aspects we don’t currently know. But aliatrary is what’s inherently uncertain. And that’s, it’s the same for throwing a dice, right? If you’re having better and better models, you’re still not going to know how, so being able to predict that with certainty is inherently uncertain. Now, at some point you get down to the subatomic level and you back to Einstein. And God does not play dice at the universe, right? So, but any practical sense, you’re never going to be able to predict the position of that pencil. And so you have this, the reason I say this is, it doesn’t mean that you can’t do it. People do it all the time. We pick up pencils. So what’s going on? What’s missing? And I think that robots and simulators have a problem because they’re very deterministic. They, the simulation has one outcome. And if you perform the same thing over robots, doing the same thing. So you tend to have a system of policy that’s trained on that particular outcome. But it’s not trained to be robust to those variations. Now, the trick to doing that is in this idea of domain randomization, right? Which is where you randomize the outputs in the simulator. But you want to do that very, very methodically. So the trick there is to have the simulator have ranges of outputs that are consistent with what you would see in simulator in reality. So this is why what we’ve been talking about recently is is real to Sim. And we learned this by through a project that we were doing on, we’re very, also working with cables. But here’s the problem is to what we call cleaner robot casting. And so here you have a robot with a weight on the end of a cable. And the robot is holding the cable above the surface. And it basically casts the cable out, like you would have fishing rod, fishing lure. And then you pick a target somewhere on the surface above where the cable has landed. And then you want the robot to do a motion that will cause the cable to wind up, but the endpoint wind up at a particular target point, meaning when it’s pulling it back or when it’s casting it out. No, when it’s pulling it back. So you want to sort of motion like this that will cause a dynamic motion towards a land in this particular point. The nice thing is you can do a lot of supervised data collection, self-supervised data collection. So the system, you have a camera overhead, you have this cable set up, and the system just basically does this all day long. So we have a data on the input control that we give to the robot and where did the actual cable land. One thing is there’s this a laitory uncertainty, and you can measure it because you’ve give it exactly the same cable motion and the endpoint lands in different places. And so we can actually draw an ellipse around that and say, this is the uncertainty that’s inherent. Even though this can run all day long, it’s still only capable of generating a few thousand examples. You really need many more to train a reliable policy. So we wanted to use a simulator. Now, what we found was that the simulator’s there are a number of simulation packages that can do this. Mojoko, Isaac, Sim, and others. And they all look good. They all kind of, when you look at them, they look very similar to what we’re seeing in the physical space. But then we try to actually give it the true parameters of what we’re measuring, and there’s a deviation. So this is where the question is, how do we tune that simulator to closely match reality? So that’s the real to Sim. So you see the difference because if you just use Mojoko and you try to have a walking machine, and you just start with the simulator, right? You get this thing and it trains and it runs, and then you say, OK, now our policy, let me go pull it on, try it on a real robot, and it doesn’t work. But if you start out by saying, I have a real robot, and now I want to make a simulator that really mirrors what’s going on with that robot. And this is closely related. Maybe it’s very similar to the idea of the digital twin. That’s very popular now. And the key is, how do you actually do this tuning systematically? And people call that system identification. Right? That’s a very common term. But there’s a lot of misperceptions about that. System ID is fairly well understood if you know the structure of the system. If you have equations that decide the system like a pendulum, and you want to identify what is the mass at the tip of that pendulum, then system ID is very good for telling you that. But if you have a system like this piece of cable and the frictional interactions of something sliding across the surface, then you don’t have a structural one. And so it’s very hard to figure out what should the parameters be in the simulation. By the way, simulation has a lot of parameters. There’s things like torsion of the cable. There’s friction of everything. There’s inner, you know, when the cable rubs against itself, there’s another frictional property, another frictional parameter for that. So there’s a dozen or more. And now you have a nice optimization problem because you have a bunch of data that you collect in real. And you want to tune the simulator to match that. So we’ve been exploring that in this context. We actually turned that in the paper, real to sim to real. We start out with real, we tune the simulator, then we can generate lots of examples. We combine that with the limited number of samples we got in real, and then train a policy, and then bring that back into real. And so, and that seems to perform better, much better, than if we just use a very limited amount of real data, which is all we can get. Or if we just use all the simulated data that wasn’t tuned to the real system. So I’ve been excited about this because I think it applies to so many problems that we’re looking at in robotics, where we look at, we have these systems, and we really want to have a simulator that’s very physical, react to it. And what’s also been very interesting, as you know, is that Nvidia has made a major push in simulation, so they’ve got a huge team of various researchers developing Isaac Sim and Variant, and also thinking about how to make those run very fast. And in parallel, deep mind acquired Mojoco last year. And that was, it was almost exactly a year ago, that was a major, major milestone because Mojoco was a very good system, but it was run by a fairly small team. Now it went into a deep mind, which has, you know, much more resources, and they’ve assembled a fantastic team of physicists and researchers to basically take Mojoco to an entirely new level of realism. So it’s been fantastic to have these two projects coming along where they’re both getting the simulators better and better. And that, I think, is going to lead to major breakthroughs in the field. You know, along these lines, I wonder if you have been involved in exploring the possible impact of causal-based models here. I’m thinking about the ellipse that you’re describing around kind of some ideal point, you know, where you’re casting back to, and you’ve got a bunch of different sources of possible, you know, call it noise. You know, you’ve got measurement noise, you’ve got control noise, various other things. You know, is there some kind of, are folks looking at causality as a way to understand how these inputs combine to create uncertainty and to create better models in the robotic realm? Well, I would have to say one answer to that is by trying to figure out how to optimize the tuning process. And that is that there is a causality inherent in controllable in the sense that the robot is able to change its parameters. And you want to be able to do that systematically. So one way to do it, let’s just take the planar robot casting is you pretty much randomly generate a lot of control inputs, trajectories for the alarm. And then you just observe where the endpoint of this cable winds up on the surface. Now, you can just generate a big data set and then throw it in and then try and analyze that to come up with a model. A better way is to do that systematically where you start doing some random examples, but then you start testing basically values of those parameters and going back out into the real system and fine tuning it so that also regions of the state space that you hadn’t explored earlier, you can, or you haven’t explored sufficiently, you can reevaluate and do more experiments in that area. So that I think is really interesting where the experiments in real are costly, if you will. They require time and an offer and it’s very difficult to reset the system, to obtain the same input, to run it again, right? But it’s very, so you want to be thoughtful and systematic and this is related to the theory of the design of experiments. And typically you’re trying to maximize some kind of mutual information gain that you will, by doing this experiments, you want to choose the experiment that’s going to gain you the most information. But it turns out computing that, solving that, is a very difficult problem. So there’s so many interesting open problems here and it’s very exciting to see how the field is maturing. I think robotics is moving at a remarkable pace, but it’s still from the public perception very far from what people commonly think should happen. And people are still thinking, well, the robot, why don’t we have our robot drivers? Why don’t we have our robots in the kitchen and robots at home taking care of us? And these things are still very far off, unfortunately. And maybe digging into that a little bit where maybe three weeks beyond Tesla’s Optimus robot announcement, which is causing people to ask the question again, oh, hey, are we close to our robot in the home? Elon says we are. What’s your take on Optimus? What was really demonstrated there, the extent to which it demonstrates that we’re close to practical, everyday robotics? OK, so I do have a take on this. I was very interested. I watched it right as it came out. And it was very interesting. I mean, I do have to hand it to Elon Musk. He’s a great entertainer. He has a real knack for doing things and doing stunts and basically having ideas that are really out there. But he’s also has been a visionary. He has actually succeeded in certain categories. So he’s done with electric cars and with Tesla with in terms of batteries, in terms of space and landing an aircraft, a rocket back on the Earth. That’s remarkable. And those are, you have to put those into context of everything else that he’s doing. And so I have to say, my first reaction is was, look, what’s going on here? He’s probably keenly aware that the price earnings ratio of an auto company is kind of at one level. But the price earnings ratio of a robotics company is much higher. So if he transforms Tesla into a robotics company, there’s a very clear benefit for that. So that could be one part of what he’s thinking. But I think what’s also going on is that he’s saying that he really is putting out a bold new idea. And he’s not afraid to take risks like that. Now, what I was happy to see was that robot was substantial progress from the past year when they had first announced it. And I remember when he very first announced it, I thought, what are you talking about, a humanoid? That’s not going to happen. But he’s really put real research behind it. Now, it was felt short of anyone’s expectation, if you know about what’s going on with agility, robotics, and Boston Dynamics. But it is a start. I think that the one thing I’m very excited about is that he understands this aspect of automation in the sense that it has to make something that’s going to run reliably and cost-effectively. So when anybody has been building humanoid over the past three decades, nobody’s really talked about the cost-effective and this and that, right? But he went out and said $20,000, right? Well, what does that mean? That means he’s going to have to develop some new motors because he’s got a lot of motors in the system. What would you say the list price is on the analog? Oh, the Alex? Oh, I haven’t seen. But it’s probably over $200,000. Order of magnitude, a couple of order of magnitude, maybe? Yes. And the thing is that these are, you know, you have to amortize all the research over the volume. So if you have, if you’re able to produce in volume, you can do this. The challenges in designing new motors, gear trains, sensors, that these are all things we actually need in robotics. You know, there’s a number of arms out there, but they’re really, they’re all still fairly expensive and either imprecise or dangerous, right? So I think there’s, I would love to see, and I think Tesla is in a perfect position to do this, is that they will come out with a new line of motors and those could be used for robots. They might even come out, here’s something I’d be saying, I think they could come out with an industrial robot arm, that Tesla would come out with a new arm, that would be, that would actually be very high precision, low cost, low mass, and we need that. So particularly if it’s one that’s based on their own needs and experience as a manufacturer. Exactly, exactly, because that’s right. So he has a use case right there, right? And there’s all kinds of aspects of the, and they tried to automate with existing robots and had a number of challenges. If they build their own robots now, that’s very, that really changes the equation. No, no one else, no other company has that big of a use case and production capability. So they could do it. And the other is sensors. We were just talking about the connect and all the lines of 3D sensors. They could really put their muscle behind a really nice, compact sensor that could be, they could give us 3D or very fast 2D sensing, and that would be very vital. Tactile sensing, by the way, that’s also heated up since the last time we talked that Facebook has really developed a partnership with, with Jelsight and come out with the digit tactile sensor. And that is very interesting, by the way, that’s a big breakthrough. And we’ve been using that, and now Jelsight just announced a new version of their sensor, the higher resolution, and they’re faster. We’re very interested in, this is like a new generation of tactile sensing that I think we’re going to see lots of applications. Okay, that’s super interesting. We may have talked about this last time. I remember as a kid taking a paeze-electric foam or something like that and slapping it between a couple of circuit boards printed on one side and using that as like a touch-ish pressure sensor. Exactly, exactly. You have a good memory. I, we did talk about that because that was where I got started as an undergrad, doing trying to build touch sensors. And it’s that, you know, you don’t have that yet. And the Jelsight is kind of another breakthrough just as connect was that it starts to make that more, more feasible, and because it’s using optics, it’s kind of riding the curve of advances in cameras. So it sounds like kind of the, the, the, to net out your take on optimists. There’s some interesting things that you hope grow out of it, but you didn’t necessarily see anything that, you know, if you were Boston dynamics would make you fear for the future of your, your own company. No, but I would have to say, I think it’s good for the field. So when you have, you know, you have someone with that level of attention and, and, and, mind share, coming out and saying robotics is where we’re going to make major advances. That is good for the whole field. I think it, it talks to young engineers who, you know, want to take a robotics class or want to maybe go into the field. It also, it speaks to investors who are, you know, look at his track record and say, hey, maybe he’s going to pull us off. And I think it would be, you know, don’t, I think a bit very unwise to bet against him. In other words, I’m not saying he’s going to come out with a, with a practical humanoid. I don’t think that’s, I think he’s going to quickly discover how complicated that is. Yeah, I think that’s what, what I’m kind of getting at or trying to get your take on it. Well, he, you know, I was, I was joking that, you know, it’s true, rocket, robotics isn’t rocket science. It’s, it’s actually, it’s much harder. And, and by that, you know, because this comes back to the things we were talking about, you know, basically doing a landing of a, of a rocket back on, you know, that stabilizing that is a beautiful control problem. But there’s, there’s only contact at the end of that. In relation, you have continuous contacts. And that is our very difficult and non-deterministic for all the reasons we talked about. So that, that problem is technically harder, my people. And so getting that right is going to require the next generation. I mean, we’re, that’s what I’m excited about, Sam, because I do feel like we’re at the point where the lot of vectors are lining up that we’re going, we’re going to see progress. And having someone like Elon and his, you know, army of supporters is, is a great thing for the field. Mm-hmm. Mm-hmm. I guess one more topic I want to take on before we wrap up. You’re the chief scientist at MB robotics. What is MB up to and how is it pushing the field forward? Well, I’ve been very impressed with the team at MB. The, since we started three years ago, the team has just been absolutely fantastic, very, very laser focused. Jeff Moller as the chief technology officer and really, the mastermind behind Dexnet, he has been leading the technical team and on the software side. And then Steve McKinley and David Ghealy have been working on the hardware side. So it’s a blend of, of, of very elegant new software and hardware that are coming together in these systems. And the, the CEO who Jim Leifer has this incredible background in logistics. So this actually all has happened since UNI talk last. Jim has been, you know, a history working in Walmart and working with, with a number of logistics companies. So he really understands the real problems. And then they, we’ve been working with the company named Pitney Boes, who is a, you know them? Yep, I worked at Pitney Boes as an undergrad for a co-op. I was doing, I was doing, I forget the term, like essentially laying out custom, custom chips. Really? Oh my God. So you know Pitney Boes. Pitney is very interesting. They’re, they’re an old company. They’ve been a old school company. Postal, postal leaders. 1920. So they just celebrated their 100 year anniversary. And what’s been interesting is, when you look at them, they’ve always been looking at technology for postage in various ways. And they, they, but they, what they do is they’re sort of behind the scenes of a lot of the postal sorting systems around the country. And so they install them for the US Postal Service, for UPS, for FedEx and others. So they really know this technology. And it’s been a pleasure to work with them because they, they’re really true engineers. They really try to solve problems. And so partnering with them has been terrific because we, we work side by side where they’ve essentially, you know, become our biggest customer. And we, we’re installed 60 of our systems over the summer, all across America. And they just invested in us. So what are the system? Is it, is it a, a hardware system? Yes. The system is hardware software. It’s called ambiSort. And what it does is it takes bins of packages and sort them into smaller bins according to zip code. So sorting is a little different than just grasping. You have to grasp the object and then scan it, determine what’s the code it is, put it on another, a gantry robot that then puts it out to, and drops it into a bin. So there’s quite a bit of hardware. In fact, it’s, it’s sort of the size of a living room. It actually fills up an 18 wheel truck. That’s, that’s the, that’s the, and what’s the form factor of the robot? Is it an arm type of form factor or? Yes. So there’s a standard six-degree freedom industrial arm at the front of it. It’s picking things up. And then there’s a gantry type robot. Like I think of an X, Y system that, that has a pivot that drops the package into the appropriate bin. So that whole system is called ambiSort. And involves lots of cameras, lots of safety features, lots of, of, of, of, of, of, fail-safe features. It’s a, it’s a big operation. It’s got thousands of parts. But that is, those are the systems we’re talking about. So each one of those can, can basically sort through hundreds of parcels per hour. And that is a big, it’s an interesting challenge because it’s very hard. This is a, can be back breaking work. Humans are prone to making a lot of mistakes. And people, the turnover in these warehouses is enormous. You know, all the companies, now Amazon is very big on this. They’re starting to, to try to find ways to automate this. And this is really our focus. Awesome. Awesome. And the, you know, for that, for, for packages, you often see, you know, as opposed to hand, you know, gripper types of actuators, like suction actuators and other things. But you mentioned DexNet, so I’m assuming you’re doing kind of more of a gripper type of actuator. Well, a great question. So DexNet 3.0 was suction, where we, we took the same, same idea and applied it to the suction model. In a sense, you have a gripper is a two point contact. You have to find two points on the object of pair. In suction, you have to find one. And so it’s one point contact. But the physics are very different. And so the resistance to shear forces, for example, are much lower for a suction cup than a gripper. So in the, the suction cup can be, is actually the, the workhorse for this kind of work in industry. And we can extend DexNet in a number of ways to make it work in this context. And that’s really been where the team has been pushing the envelope. And we also collect data from every one of these systems. So it’s a, it’s a wonderful problem from a machine learning point of view. Because we have data sets. We have images. We have sensor values. We have all this stuff. We can characterize every single failure. And analyze it. And then try back testing different algorithms to be able to produce those. And that’s where there’s a huge opportunity. Because really, there’s a gap. And can we start closing it to really increase the throughput, the picks per hour? Awesome. Awesome. Well, Ken, I think we covered a ton of ground, but also demonstrated that it’s really hard to stick to in a half years of robotics, advancement, and innovation in an hour. So I think that just means we’ll have to be sure to catch up more frequently in the future. I would love that. Thanks, Sam. I have to say, I’ve been such a pleasure because I listened to your podcast on my bike rides, Mountain biking. And so I’ve just enjoyed them so many good hours on a bike with you. And last thing I want to say, I don’t know if that witness will air, but the conference on robot learning is going to be in New Zealand this December, 14 through the 18. And it is. So I’m chairing the conference. And we’ve been, it’s been a real pleasure. We have 500 papers submitted. A top notch group of about 200 papers will be presented there. And you can register as an online, online, to watch all the talks and everything else. For I think it’s like close to $200, not very expensive. And then I will tell your audience that we’re also going to make all this available offline after the conference. Awesome. Thank you so much, Sam. I really appreciate it. Great time to meet you. Thanks so much, Ken. All right. Take care.