Pieter Abbeel: Deep Reinforcement Learning

Pieter Abbeel is a professor at UC Berkeley, director of the Berkeley Robot Learning Lab, and is one of the top researchers in the world working on how to make robots understand and interact with the world around them, especially through imitation and deep reinforcement learning. Video version is available on YouTube https://www.youtube.com/watch?v=l-mYLq6eZPY. If you would like to get more information about this podcast go to https://lexfridman.com/ai or connect with @lexfridman on Twitter https://twitter.com/lexfridman, LinkedIn https://www.linkedin.com/in/lexfridman/, Facebook https://www.facebook.com/lexfridman, or YouTube https://www.youtube.com/lexfridman where you can watch the video versions of these conversations.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Lex Fridman?

  The following is a conversation with Peter Abbeel. He's a professor at UC Berkeley and the director of the Berkeley Robotics Learning Lab. He's one of the top researchers in the world, working on how we make robots understand and interact with the world around them, especially using imitation and deep reinforcement learning. This conversation is part of the MIT course on artificial general intelligence and the Artificial Intelligence Podcast. If you enjoy it, please subscribe on YouTube, iTunes, or your podcast provider of choice, or simply connect with me on Twitter at LexFriedman, spelled F R I D. And now, here's my conversation with Peter Abbeel.

• 0:58Lex Fridman?

  You've mentioned that if there was one person you could meet, it would be Roger Federer. So let me ask, when do you think we'll have a robot that fully autonomously can beat Roger Federer at tennis? At Roger Federer level player at tennis?

• 1:14Speaker 01

  Well, first, if you can make it happen for me to meet Roger, let me know. In terms of getting a robot to beat him at tennis, it's kind of an interesting question because for a lot of the challenges we think about in AI, the software is really the missing piece. But for something like this, the hardware is nowhere near either. Like, to really have a robot that can physically run around, the Boston Dynamics robots are starting to get there, but still not really human-level ability to run around and then swing a racket.

• 1:53Lex Fridman?

  So you think that's a hardware problem?

• 1:55Speaker 01

  I don't think it's a hardware problem only. I think it's a hardware and a software problem. I think it's both. And I think they'll have independent progress. So I'd say the hardware maybe in 10, 15 years.

• 2:08Lex Fridman?

  On clay, not grass. I mean, grass is probably harder.

• 2:11Speaker 01

  With the sliding? Well, the clay, I'm not sure what's harder, grass or clay. The clay involves sliding, which might be harder to master, actually. Yeah.

• 2:22Lex Fridman?

  But you're not limited to bipedal. I mean, I'm sure there's no-

• 2:26Speaker 01

  Well, if we can build a machine, it's a whole different question, of course. If you can say, okay, this robot can be on wheels. It can move around on wheels and can be designed differently. Then I think that can be done sooner probably than a full humanoid type of setup.

• 2:43Lex Fridman?

  What do you think of swing a racket? So you've worked at basic manipulation. How hard do you think is the task of swinging a racket? Would be able to hit a nice backhand or a forehand?

• 2:56Speaker 01

  Let's say we just set up stationary, a nice robot arm, let's say. You know, a standard industrial arm. And it can watch the ball come and then swing the racket. It's a good question. I'm not sure it would be super hard to do. I mean, I'm sure it would require a lot. If we do it with reinforcement learning, it would require a lot of trial and error. It's not going to swing it right the first time around. But yeah, I don't see why it couldn't.

• 3:24Lex Fridman?

  So you think it's learnable?

• 3:25Speaker 01

  Swing it the right way. I think it's learnable. I think if you set up a ball machine, let's say on one side, and then a robot with a tennis racket on the other side, I think it's learnable. And maybe a little bit of pre-training and simulation? Yeah, I think that's feasible. I think the swing the racket is feasible. It'd be very interesting to see how much precision it can get.

• 3:48Speaker 01

  Because I mean, that's where, I mean, some of the human players can hit it on the lines, which is very high precision.

• 3:56Lex Fridman?

  With spin. The spin is an interesting, whether RL can learn to put a spin on the ball.

• 4:02Speaker 01

  Well, you got me interested. Maybe someday we'll set this up.

• 4:05Lex Fridman?

  Sure.

• 4:06Speaker 01

  You got me intrigued.

• 4:07Lex Fridman?

  Your answer is basically, okay, for this problem, it sounds fascinating, but for the general problem of a tennis player, we might be a little bit farther away. What's the most impressive thing you've seen a robot do in the physical world?

• 4:20Speaker 01

  So physically, for me, it's the Boston Dynamics videos always just bring home and just super impressed. Recently, the robot running up the stairs, doing the parkour type thing. I mean, yes, we don't know what's underneath. They don't really write a lot of detail, but even if it's hard coded underneath, which it might or might not be, just the physical abilities of doing that parkour, that's a very impressive robot right there.

• 4:50Lex Fridman?

  So have you met Spot Mini or any of those robots in person?

• 4:53Speaker 01

  Met Spot Mini last year in April at the MARS event that Jeff Bezos organizes. They brought it out there and it was nicely following around Jeff. When Jeff left the room, they had it follow him along, which is pretty impressive.

• 5:08Lex Fridman?

  So I think there's some confidence to know that there's no learning going on on those robots. The psychology of it, so while knowing that, while knowing there's not, if there's any learning going on, it's very limited. I met Spot Mini earlier this year and knowing everything that's going on, having one-on-one interactions, so I get to spend some time alone. And there's immediately a deep connection on the psychological level. Even though you know the fundamentals, how it works, there's something magical. So do you think about the psychology of interacting with robots in the physical world? Even you just showed me the PR2, the robot, and there was a little bit, something like

• 5:52Lex Fridman?

  a face, had a little bit, something like a face, there's something that immediately draws you to it. Do you think about that aspect of the robotics problem?

• 6:01Speaker 01

  Well, it's very hard with Brett here. We'll give him a name, Berkeley Robot for the Elimination of Tedious Tasks. It's very hard to not think of. The robot as a person and it seems like everybody calls him a he for whatever reason, but that also makes it more a person than if it was an it. And it's, it seems pretty natural to think of it that way. This past weekend really struck me. I've, I've seen Pepper many times on, on videos. But then I was at an event organized by, this was by Fidelity and they had scripted Pepper to help moderate some sessions. And they had scripted Pepper to have the personality of a child a little bit.

• 6:43Speaker 01

  And it was very hard to not think of it as its own person in some sense, because it was just gonna jump, it would just jump into conversation, making it very interactive. Moderator would be saying, Pepper would just jump in, "Hold on, how about me? Can I participate in this too?" And just like, okay, this is like, like a person. And that was 100% scripted. And even then, it was hard not to have that sense of somehow there is something there.

• 7:06Lex Fridman?

  So as we have robots interact in this physical world, is that a signal that could be used in reinforcement learning? You've, you've worked a little bit in this direction, but do you think that, that psychology can be somehow pulled in?

• 7:19Speaker 01

  Yes, that's a question, I would say, a lot, a lot of people ask. And I think part of why they ask it is they're thinking about, how unique are we really still as people? Like after they see some results, they see a computer play, go, they say, "Computer do this, that." They're like, okay, but can it really have emotion? Can it really interact with us in that way? And then, once you're around robots, you already start feeling it. And I think that kind of maybe methodologically, the way that I think of it is, if you run something like reinforcement learning, it's about optimizing some objective. And there's no reason that the objective couldn't be tied into

• 8:04Speaker 01

  How much does a person like interacting with this system? And why could not the reinforcement learning system optimize for the robot being fun to be around? And why wouldn't it then naturally become more and more attractive and more and more, maybe like a person or like a pet? I don't know what it would exactly be, but more and more have those features and acquire them automatically?

• 8:25Lex Fridman?

  As long as you can formalize an objective of what it means to like something. What, how you exhibit, what's the ground truth? How do you, how do you get the reward from human? Because you have to somehow collect that information from your human. But you're saying if you can formulate as an objective, it can be learned.

• 8:44Speaker 01

  There is no reason it couldn't emerge through learning. And maybe one way to formulate as an objective, you wouldn't have to necessarily score it explicitly. So standard rewards are numbers, and numbers are hard to come by. This is a 1.5 or 1.7 on some scale. It's very hard to do for a person. But much easier is for a person to say, okay, what you did the last five minutes was much nicer than what you did the previous five minutes. And that now gives a comparison. And in fact, there have been some results on that. For example, Paul Christiano and collaborators at OpenAI had the Hopper, Mojoko Hopper, a one-legged robot.

• 9:18Lex Fridman?

  The backflip.

• 9:19Speaker 01

  Learn to do backflips. Purely from feedback, I like this better than that. That's kind of equally good. And after a bunch of interactions, it figured out what it was the person was asking for, namely a backflip. And so I think the same thing.

• 9:32Lex Fridman?

  It wasn't trying to do a backflip. It was just getting a score from the comparison score from the person based on.

• 9:40Speaker 01

  Person having in mind, in their own mind, I wanted to do a backflip. But the robot didn't know what it was supposed to be doing. It just knew that sometimes the person said, this is better, this is worse. And then the robot figured out what the person was actually after was a backflip. And I'd imagine the same would be true for things like more interactive robots, that the robot would figure out over time, oh, this kind of thing apparently is appreciated more than this other kind of thing.

• 10:07Lex Fridman?

  So when I first picked up Sutton's, Richard Sutton's reinforcement learning book, before sort of this deep learning, before the reemergence of neural networks as a powerful mechanism for machine learning, RL seemed to me like magic. It was beautiful. So that seemed like what intelligence is, RL reinforcement learning. So how do you think we can possibly learn anything about the world when the reward for the actions is delayed, is so sparse? Like where is, why do you think RL works? Why do you think you can learn anything under

• 10:50Lex Fridman?

  such sparse rewards, whether it's regular reinforcement learning or deep reinforcement learning? What's your intuition?

• 10:56Speaker 01

  The counterpart of that is, why is RL, why does it need so many samples, so many experiences to learn from? Because really what's happening is when you have a sparse reward, you do something, maybe for like, I don't know, you take 100 actions and then you get a reward. And maybe you get like a score of three. And I'm like, okay, three, not sure what that means. You go again and now you get two. And now you know that that sequence of 100 actions that you did the second time around somehow was worse than the sequence of 100 actions you did the first time around. But that's tough to now know which one of those were better or worse. Some might have been good and bad in either one. And so that's why you need so many experiences. But once you have enough experiences, effectively

• 11:38Speaker 01

  RL is teasing that apart. It's trying to say, okay, what is consistently there when you get a higher reward and what's consistently there when you get a lower reward? And then kind of the magic of sometimes the policy gradient update is to say, now let's update the neural network to make the actions that were kind of present when things are good more likely, and make the actions that are present when things are not as good, less likely.

• 12:01Lex Fridman?

  So that's, that is the counterpoint, but it seems like you would need to run it a lot more than you do, even though right now people could say that RL is very inefficient. But it seems to be way more efficient than one would imagine on paper, that the simple updates to the policy, the policy gradient, that somehow you can learn, exactly as you just said, what are the common actions that seem to produce some good results, that that somehow can learn anything? It seems counterintuitive at least. Is there some intuition behind that?

• 12:33Speaker 01

  Yeah, so I think There's a few ways to think about this. The way I tended to think about it mostly originally when, so when we started working on deep reinforcement learning here at Berkeley, which was maybe 2011, 12, 13, around that time, John Schulman was a PhD student initially kind of driving it forward here. And the way we thought about it at the time was if you think about rectified linear units or kind of rectifier type neural networks, what do you get? You get something that's piecewise linear feedback control. And if you look at the literature, linear feedback control is extremely successful,

• 13:16Speaker 01

  can solve many, many problems surprisingly well. I remember, for example, when we did helicopter flight, if you're in a stationary flight regime, not a non-stationary, but a stationary flight regime like hover, you can use linear feedback control to stabilize a helicopter, a very complex dynamical system, but the controller is relatively simple. And so I think that's a big part of is that if you do feedback control, even though the system you control can be very, very complex, often relatively simple control architectures can already do a lot, but then also just linear is not good enough. And so one way you can think of these neural networks is that in some ways they tile the space, which people were already trying to do more by hand or with finite state machines, say this linear controller here, this linear controller here.

• 14:00Speaker 01

  Neural network learns to tile the space and say linear controller here, another linear controller here, but it's more subtle than that.

• 14:06Lex Fridman?

  Yeah.

• 14:06Speaker 01

  And so it's benefiting from this linear control aspect, it's benefiting from the tiling, but it's somehow tiling it one dimension at a time. Because if, let's say you have a two-layer network, even that hidden layer, you make a transition from active to inactive or the other way around, that is essentially one axis, but not axis aligned, but one direction that you change. And so you have this kind of very gradual tiling of the space where you have a lot of sharing between the linear controllers that tile the space. And that was always my intuition as to why to expect that this might work. Pretty well. It's essentially leveraging the fact that linear feedback control is so good, but of course not enough. And this is a gradual tiling of the space with linear feedback controls that share a lot of

• 14:51Speaker 01

  expertise across them.

• 14:53Lex Fridman?

  So that's, that's really nice intuition. But do you think that scales to the more and more general problems of when you start going up the number of dimensions? When you start going down in terms of how often you get a clean reward signal? Does that intuition carry forward to those crazier, weirder worlds that we think of as the real world?

• 15:20Speaker 01

  So I think where things get really tricky in the real world compared to the things we've looked at so far with great success in reinforcement learning is the time scales, which takes us to an extreme. So when you think about the real world, I mean, I don't know, maybe some student decided to do a PhD here, right? Okay, that's, that's a decision. That's a very high level decision. But if you think about their lives, I mean, any person's life, it's a sequence of muscle fiber contractions and relaxations, and that's how you interact with the world. And that's a very high frequency control thing, but it's ultimately what you do and how you affect the world.

• 16:03Speaker 01

  Until, I guess, we have brain readings and you can maybe do it slightly differently. But typically, that's how you affect the world. And the decision of doing a PhD is like so abstract relative to what you're actually doing in the world. And I think that's where credit assignment becomes just completely beyond what any current RL algorithm can do. And we need hierarchical reasoning at a level that is just not available at all yet.

• 16:29Lex Fridman?

  Where do you think we can pick up hierarchical reasoning? By which mechanisms?

• 16:33Speaker 01

  Yeah, so maybe let me highlight what I think the limitations are of what already was done. 20, 30 years ago. In fact, you'll find reasoning systems that reason over relatively long horizons. But the problem is that they were not grounded in the real world. So people would have to hand-design some kind of logical, dynamical descriptions of the world, and that didn't tie into perception. And so that didn't tie into real objects and so forth. And so that was a big gap. Now with deep learning, we start having the ability to really see with sensors, process that, and understand what's in the world.

• 17:18Speaker 01

  And so it's a good time to try to bring these things together. I see a few ways of getting there. One way to get there would be to say, deep learning can get bolted on somehow to some of these more traditional approaches. Now, bolted on would probably mean you need to do some kind of end-to-end training, where you say, my deep learning processing somehow leads to a representation that in turn uses some kind of traditional, underlying, dynamical systems that can be used for planning. And that's, for example, the direction Aviv Tammar and Thanar Kuratich here have been pushing with causal info again, and of course, other people too. That's one way. Can we somehow force it into the form factor that is amenable to reasoning?

• 18:02Speaker 01

  Another direction we've been thinking about for a long time and didn't make any progress on was more information theoretic approaches. So the idea there was that what it means to take high-level action is to take and choose a latent variable now that tells you a lot about what's going to be the case in the future. Because that's what it means to take a high-level action. I say, okay, I decide I'm going to navigate to the gas station because I need to get gas for my car. Well, that'll now take five minutes to get there, but the fact that I get there, I could already tell that from the high-level action I took much earlier. That, we had a very hard time getting success with.

• 18:44Speaker 01

  I'm not saying it's a dead end necessarily, but we had a lot of trouble getting that to work. And then we start revisiting the notion of what are we really trying to achieve? What we're trying to achieve is not necessarily hierarchy per se, but you could think about what does hierarchy give us? What it's, we hope it would give us is better credit assignment. Kind of what is better credit assignment is giving us, it gives us faster learning. Right? And so faster learning is ultimately maybe what we're after. And so that's what we ended up with the RL squared paper on learning to reinforcement learn, which at a time, Rocky Dwan led. And that's exactly the meta learning approach where you say, okay, we don't know how to

• 19:29Speaker 01

  design hierarchy. We know what we want to get from it. Let's just end to end optimize for what we want to get from it and see if it might emerge. And we saw things emerge. The maze navigation had consistent motion down hallways, which is what you want. A hierarchical controller should say, I want to go down this hallway. And then when there is an option to take a turn, I can decide whether to take a turn or not and repeat. Even had the notion of where have you been before or not? To not revisit places you've been before. It still didn't scale yet to the real world kind of scenarios I think you had in mind, but it was some sign of life that maybe you can meta learn these hierarchical concepts.

• 20:07Lex Fridman?

  I mean, it seems like through these meta learning concepts, we get at the, what I think is one of the hardest and most important problems of AI, which is transfer learning. So it's generalization. How far along this journey towards building general systems are we? Being able to do transfer learning well? So there's some signs that you can generalize a little bit, but do you think we're on the right path or totally different breakthroughs are needed to be able to transfer knowledge between different learned models?

• 20:47Speaker 01

  Yeah, I'm pretty torn on this in that I think there are some very impressive.

• 20:52Lex Fridman?

  The day.

• 20:53Speaker 01

  There's just some very impressive results already, right?

• 20:57Lex Fridman?

  Yes.

• 20:58Speaker 01

  I would say when, even with the initial kind of big breakthrough in 2012 with AlexNet, the initial thing is, okay, great. This does better on ImageNet, hence image recognition. But then immediately thereafter, there was of course the notion that, wow, what was learned on ImageNet and you now want to solve a new task, you can fine tune AlexNet for new tasks. And that was often found to be the even bigger deal, that you learned something that was reusable, which was not often the case before. Usually in machine learning, you learn something for one scenario and that was it.

• 21:35Lex Fridman?

  And that's really exciting. I mean, that's a huge application. That's probably the biggest success of transfer learning to date, in terms of scope and impact.

• 21:44Speaker 01

  That was a huge breakthrough. And then recently, I feel like similar kind of, by scaling things up, it seems like this has been expanded upon. People training even bigger networks, they might transfer even better. If you looked at, for example, some of the OpenAI results on language models and some of the recent Google results on language models, they're learned for just prediction and then they get reused for other tasks. And so I think there is something there where somehow if you train a big enough model on enough things, it seems to transfer. Some DeepMind results that I thought were very impressive, the Unreal results, where

• 22:25Speaker 01

  it was learned to navigate mazes in ways where it wasn't just doing reinforcement learning, but it had other objectives it was optimizing for. So I think there's a lot of interesting results already. I think. Maybe where it's hard to wrap my head around this, to which extent or when do we call something generalization, right, or the levels of generalization involved in these different tasks, right? So.

• 22:52Lex Fridman?

  You draw this, by the way, just to frame things, I've heard you say somewhere, it's the difference between learning to master versus learning to generalize, that it's a nice line to think about, and I guess you're saying it's a gray area of what learning to master and learning to generalize, where one starts.

• 23:11Speaker 01

  I think I might have heard this, I might have heard it somewhere else, and I think it might have been one of your interviews, maybe the one with Yoshua Benjamin, I'm not 100% sure, but I liked the example, and I'm gonna, not sure who it was, but the example was essentially if you use current deep learning techniques, what we're doing to predict, let's say, the relative motion of our planets, it would do pretty well, but then now if a massive new mass enters our solar system, it would probably not predict what will happen, right? And that's a different kind of generalization, that's a generalization that relies on the

• 23:52Speaker 01

  ultimate simplest, simplest explanation that we have available today to explain the motion of planets, whereas just pattern recognition could predict our current solar system motion pretty well, no problem. And so I think that's an example of a kind of generalization that is a little different from what we've achieved so far, and it's not clear if just, you know, regularizing more and forcing it to come up with a simpler, simpler, simpler explanation, say, look, this is not simple, but that's what physics researchers do, right, they say, can I make this even simpler? How simple can I get this? What's the simplest equation that can explain everything, right? The master equation for the entire dynamics of the universe, we haven't really pushed that direction as hard in deep learning, I would say.

• 24:37Speaker 01

  Not sure if it should be pushed, but it seems a kind of generalization you get from that that you don't get in our current methods so far.

• 24:43Lex Fridman?

  So I just talked to Vladimir Vapnik, for example, who is a statistician, statistical learning, and he kind of dreams of creating the E equals mc squared for learning, right, the general theory of learning. Do you think that's a fruitless pursuit?

• 25:03Lex Fridman?

  In the near term, within the next several decades?

• 25:08Speaker 01

  I think that's a really interesting pursuit. And in the following sense, in that there is a lot of evidence that

• 25:18Speaker 01

  the brain is pretty modular. And so I wouldn't maybe think of it as the theory, maybe the underlying theory, but more kind of the principle. Where there have been findings where

• 25:31Speaker 01

  people who are blind will use the part of the brain usually used for vision, for other functions. And even after some kind of, if people get rewired in some way, they might be able to reuse parts of their brain for other functions. And so what that suggests is some kind of modularity. And I think it is a pretty natural thing to strive for to see, can we find that modularity? Can we find this thing? Of course, it's not every, every part of the brain is not exactly the same. Not everything can be rewired arbitrarily. But if you think of things like the neocortex, which is a pretty big part of the brain, that seems fairly modular from what the findings so far. Can you design something equally modular?

• 26:15Speaker 01

  And if you can just grow it, it becomes more capable, probably. I think that would be the kind of interesting underlying principle to shoot for that is not unrealistic.

• 26:26Lex Fridman?

  Do you think you prefer math or empirical trial and error for the discovery of the essence of what it means to do something intelligent? So reinforcement learning in bodies, both groups, right? To prove that something converges, prove the bounds. And then at the same time, a lot of the successes are, well, let's try this and see if it works. So which do you gravitate towards? How do you think of those two parts of your brain?

• 26:53Speaker 01

  So?

• 26:56Speaker 01

  Maybe I would prefer we could make the progress with mathematics. And the reason maybe I would prefer that is because often if you have something you can mathematically formalize, you can leapfrog a lot of experimentation. And experimentation takes a long time to get through and a lot of trial and error, kind of reinforcement in your research process. But you need to do a lot of trial and error before you get to a success. So if you can leapfrog that, to my mind, that's what the math is about. And hopefully, once you do a bunch of experiments, you start seeing a pattern. You can do some derivations that leapfrog some experiments. But I agree with you. I mean, in practice, a lot of the progress has been such that we have not been able to find the math that allows us to leapfrog ahead.

• 27:41Speaker 01

  And we are kind of making gradual progress, one step at a time. A new experiment here, a new experiment there that gives us new insights and gradually building up, but not getting to something yet where we're just, okay, here is an equation that now explains how, you know, that would be, have been two years of experimentation to get there, but this tells us what the result's going to be. Unfortunately, not so much yet.

• 28:04Lex Fridman?

  Not so much yet, but your hope is there. In trying to teach robots or systems to do everyday tasks, or even in simulation, what do you think you're more excited about? Imitation learning or self-play? So letting robots learn from humans? Or letting robots play on their own, try to figure out in their own way and eventually play, eventually interact with humans or solve whatever problem is? What's the more exciting to you? What's more promising, you think, as a research direction?

• 28:42Speaker 01

  So when we look at...

• 28:47Speaker 01

  Self-play. What's so beautiful about it is it goes back to the challenges in reinforcement learning. So the challenge of reinforcement learning is getting signal. And if you never succeed, you don't get any signal. In self-play, you're on both sides, so one of you succeeds. And the beauty is also one of you fails. And so you see the contrast. You see the one version of me that did better than the other version. And so every time you play yourself, you get signal. And so whenever you can turn something into self-play, you're in a beautiful situation where you can naturally learn much more quickly than in most other reinforcement learning environments. So I think if somehow we can turn more reinforcement learning problems into self-play formulations,

• 29:32Speaker 01

  that would go really, really far. So far, self-play has been largely around games where there is natural opponents. But if we could do self-play for other things, and let's say, I don't know, a robot learns to build a house. I mean, that's a pretty advanced thing to try to do for a robot, but maybe it tries to build a hut or something. If that can be done through self-play, it would learn a lot more quickly if somebody can figure it out. And I think that would be something where it goes closer to the mathematical leapfrogging, where somebody figures out a formalism to say, okay, any RL problem by playing this and this idea, you can turn it into a self-play problem where you get signal a lot more easily.

• 30:09Speaker 01

  Reality is, many problems we don't know how to turn to self-play. And so either we need to provide detailed reward, that doesn't just reward for achieving a goal, but rewards for making progress, and that becomes time-consuming. And once you're starting to do that, let's say you want a robot to do something, you need to give all this detailed reward, well, why not just give a demonstration? Because why not just show the robot? And now the question is, how do you show the robot? One way to show is to tele-operate a robot, and then the robot really experiences things. And that's nice, because that's really high signal to noise ratio data. And we've done a lot of that. And you teach a robot skills, in just 10 minutes, you can teach a robot a new basic skill, like, okay, pick up the bottle, place it somewhere else. That's a skill, no matter where the bottle starts, maybe it always goes onto a target or something.

• 30:53Speaker 01

  That's fairly easy to teach a robot with tele-op. Now... What's even more interesting if you can now teach a robot through third-person learning, where the robot watches you do something and doesn't experience it, but just watches it and says, "Okay, well, if you're showing me that, that means I should be doing this." And I'm not going to be using your hand, because I don't get to control your hand, but I'm going to use my hand. I do that mapping. And so that's where I think one of the big breakthroughs has happened this year. This was led by Chelsea Finn here. It's almost like learning a machine translation for demonstrations, where you have a human demonstration and the robot learns to translate it into what it means for the robot to do it. And that was a meta-learning formulation, learn from one to get the other.

• 31:37Speaker 01

  And that, I think, opens up a lot of opportunities to learn a lot more quickly.

• 31:41Lex Fridman?

  So my focus is on autonomous vehicles. Do you think this approach of third-person watching, the autonomous driving is amenable to this kind of approach?

• 31:50Speaker 01

  So for autonomous driving, I would say it's... Third-person is slightly easier. And the reason I'm going to say it's slightly easier to do with third-person is because the car dynamics are very well-understood. So the...

• 32:07Lex Fridman?

  Easier than first-person, you mean?

• 32:10Speaker 01

  I think... So I think the distinction between third-person and first-person is not a very important distinction for autonomous driving. They're very similar. Because the distinction is really about who turns the steering wheel. And... Or maybe, I'll maybe put it differently. How to get from a point where you are now to a point, let's say, a couple meters in front of you. And that's a problem that's very well-understood. And that's the only distinction between third and first-person there. Whereas with the robot manipulation, interaction forces are very complex, and it's still a very different thing. For autonomous driving, I think there is still the question imitation. Versus RL. So imitation gives you a lot more signal.

• 32:53Speaker 01

  I think where imitation is lacking and needs some extra machinery is it doesn't, in its normal format, doesn't think about goals or objectives. And of course, there are versions of imitation learning, inverse reinforced learning type imitation learning, which also thinks about goals. I think then we're getting much closer. But I think it's very hard to think of a fully reactive car generalizing well. If it really doesn't have a notion of objectives to generalize well to the kind of general that you would want. You'd want more than just that reactivity that you get from just behavioral cloning slash supervised learning.

• 33:33Lex Fridman?

  So a lot of the work, whether it's self-play or even imitation learning, would benefit significantly from simulation, from effective simulation. And you're doing a lot of stuff in the physical world and in simulation. Do you have hope for greater and greater power of simulation being boundless eventually? To where most of what we need to operate in the physical world would could be simulated to a degree that's directly transferable to the physical world? Or are we still very far away from that?

• 34:08Speaker 01

  So I think we could even rephrase that question in some sense.

• 34:15Lex Fridman?

  Please.

• 34:17Speaker 01

  And so, so the power of simulation, right? As simulators get better and better, of course, becomes stronger and we can learn more in simulation. But there's also another version, which is where you say the simulator doesn't even have to be that precise, as long as it's somewhat representative. And instead of trying to get one simulator that is sufficiently precise to learn in and transfer really well to the real world, I'm going to build many simulators.

• 34:43Lex Fridman?

  Ensemble of simulators?

• 34:45Speaker 01

  Ensemble of simulators. Not any single one of them is sufficiently representative of the real world, such that it would work if you train in there. But if you train in all of them. Then there is something that's good in all of them. The real world will just be, you know, another one of them. That's, you know, not identical to any one of them, but just another one of them.

• 35:07Lex Fridman?

  Another sample from the distribution of simulators.

• 35:09Speaker 01

  Exactly.

• 35:10Lex Fridman?

  We do live in a simulation, so this is just one, one other one.

• 35:14Speaker 01

  I'm not sure about that, but yeah.

• 35:18Speaker 01

  It's definitely a very advanced simulator if it is.

• 35:20Lex Fridman?

  Yeah, it's a pretty good one. I've talked to Russell, it's something you think about a little bit too. Of course, you're like really trying to build these systems, but do you think about the future of AI? A lot of people have concern about safety. How do you think about AI safety as you build robots that are operating in the physical world? What is, yeah, how do you approach this problem in an engineering kind of way, in a systematic way?

• 35:46Speaker 01

  So when a robot is doing things, you kind of have a few notions of safety to worry about. One is that the robot is physically strong and of course could do a lot of damage. Same for cars, which we can think of as robots do in some way. And this could be completely unintentional. So it could be not the kind of long-term AI safety concerns that, okay, AI is smarter than us. And now what do we do? But it could be just very practical. Okay, this robot, if it makes a mistake, what are the results going to be? Of course, simulation comes in a lot there to, to test in simulation.

• 36:23Speaker 01

  It's a difficult question and I'm always wondering, like I always wonder, let's say you look at, let's go back to driving, because a lot of people know driving well, of course. What do we do to test somebody for driving, right? To get a driver's license? What do they really do? I mean, you fill out some test and then you drive and I mean,

• 36:44Lex Fridman?

  For a few minutes.

• 36:44Speaker 01

  It's suburban California. That driving test is just you drive around the block, pull over. You do a stop sign successfully and then, you know, you pull over again and you're pretty much done. And you're like, okay, if a self-driving car did that, would you trust it that it can drive? I'd be like, no, that's not enough for me to trust it. But somehow for humans, we've figured out that somebody being able to do that is representative of them being able to do a lot of other things. And so I think somehow for humans, we figured out representative tests of what it means if you can do this, what you can really do. Of course, testing humans, humans don't want to be tested at all times. Self-driving cars or robots could be tested more often, probably.

• 37:28Speaker 01

  You can have replicas that get tested and they're known to be identical because they use the same neural net and so forth. But still, I feel like we don't have this kind of unit tests or proper tests for, for robots. And I think there's something very interesting to be thought about there, especially as you update things. Your software improves, you have a better self-driving car suite, you update it. How do you know it's indeed more capable on everything than what you had before, that you didn't have any bad things creep into it? So I think that's a very interesting direction of research, that there is no real solution yet, except that somehow for humans we do. Because we say, okay, you have a driving test, you passed, you can go on the road now.

• 38:08Speaker 01

  And humans have accidents every like million or 10 million miles, something pretty phenomenal compared to that short test that is being done.

• 38:18Lex Fridman?

  So let me ask, you've mentioned, you've mentioned that Andrew Ng, by example, showed you the value of kindness. Do you think the space of policies, good policies for humans and for AI is populated by policies that with kindness or ones that are the opposite? Exploitation, even evil? So if you just look at the sea of policies we operate under as human beings, or if AI system had to operate in this real world, do you think it's really easy to find policies that are full of kindness, like we naturally fall into them? Or is it like a very hard optimization problem?

• 39:04Speaker 01

  I mean, there is kind of two optimizations happening for humans, right? So for humans, there's kind of the very long-term optimization, which evolution has done for us. And we're kind of predisposed to like certain things. And that's in some sense what makes our learning easier, because, I mean, we know things like pain and hunger and thirst. And the fact that we know about those is not something that we were taught. That's kind of innate. When we're hungry, we're unhappy. When we're thirsty, we're unhappy. When we have pain, we're unhappy. And ultimately, evolution built that into us to think about those things. And so I think there is a notion that it seems somehow humans evolved in general

• 39:45Speaker 01

  to prefer to get along in some ways, but at the same time also to be very territorial and kind of centric to their own tribe. Like, it seems like that's the kind of space we converge onto. I mean, I'm not an expert in anthropology, but it seems like we're very kind of good within our own tribe, but need to be taught to be nice to other tribes.

• 40:11Lex Fridman?

  Well, if you look at Steven Pinker, he highlights this pretty nicely in Better Angels of Our Nature, where he talks about violence decreasing over time consistently. So whatever tension, whatever teams we pick, it seems that the long arc of history goes towards us getting along more and more.

• 40:31Speaker 01

  I hope so.

• 40:33Lex Fridman?

  So do you think that... Do you think it's possible to teach RL-based robots this kind of kindness? This kind of ability to interact with humans? This kind of policy? Even to... Let me ask, let me ask a fun one. Do you think it's possible to teach RL-based robot to love a human being and to inspire that human to love the robot back? So to like a RL-based algorithm that leads to a happy marriage?

• 41:04Speaker 01

  That's an interesting question. Maybe I'll answer it with another question, right?

• 41:11Speaker 01

  Because I mean, but I'll come back to it. So another question you can have is, okay, I mean, how close does some people's happiness get from interacting with just a really nice dog? Like, I mean, dogs, you come home, that's what dogs do. They greet you. They're excited. Makes you happy when you come home to your dog. You're just like, okay, this is exciting. They're always happy when I'm here. And if they don't greet you, because maybe, whatever, your partner took them on a trip or something, you might not be nearly as happy when you get home, right? And so the kind of, it seems like the level of reasoning a dog has is pretty sophisticated, but then is still not yet at the level of human reasoning. And so it seems like we don't even

• 41:53Speaker 01

  need to achieve human level reasoning to get like very strong affection with humans. And so my thinking is, why not, right? Why couldn't, with an AI, couldn't we achieve the kind of level of affection that humans feel among each other or with friendly animals and so forth?

• 42:14Speaker 01

  So question, is it a good thing for us or not? That's another thing, right? Because, I mean,

• 42:20Speaker 01

  but I don't see why not.

• 42:22Lex Fridman?

  Why not? Yeah. So Elon Musk says love is the answer. Maybe he should say, love is the objective function, and then RL is the answer, right?

• 42:32Speaker 01

  Well, maybe.

• 42:34Lex Fridman?

  Oh, Peter, thank you so much. I don't want to take up more of your time. Thank you so much for talking today.

• 42:38Speaker 01

  Well, Lex, thanks for coming by. Great to have you visit.