01:00:29 · Mario Gabriele, Karol Hausman
[01:00:29] S2: I'm quite open minded about these things. I'm not [01:00:31] S0: very pragmatic about the path that's gonna it's gonna take to get there. I think simulation over the past, you know, a decade plus has been improving like crazy. Simulation that we see today is much more realistic than what we've seen in the past and it's becoming more and more scalable. I believe that the first place where we're gonna see the impact of it is not going to be as much data collection as it's going to be evaluation. And one thing we see here at at at Physical Intelligence is that as these models become more powerful, it takes more and more time to evaluate them. So now not only you need to evaluate them for longer, you need to take more samples to distinguish between the model being, you know, at 99% success rate versus 95% rate, the success rate, and it was between, you know, 50 versus 70%. You just need more samples to be statistically significant, but the repertoire becomes broader and broader. So now you need to evaluate them on more tasks across more robots in more environments. And that trend, I think, will continue. As the models get get stronger, you will need to evaluate them for much longer and much more sophisticated ways. So I think if simulation starts to work in the way that we all hope it will, I think that's where it will... We'll see the the first impact of it, where we can start evaluating these models in simulation across diverse set of scenarios. And if they're realistic enough, that should correlate with with real world evaluations.
01:01:58 · Mario Gabriele
[01:01:58] S1: You know, we talked about some of the real inflection points for you over the course of your career. I'm sure that physical intelligence biggest inflection points are are to come. But over the past couple of years so far, what have been the... Maybe the biggest surprises or the the moments that have impressed you the most?
01:02:17 · Karol Hausman
[01:02:17] S0: I think the the main thing is just how fast all of this has been moving and how much it's picking up speed as we go. So before we started the company, we thought that it was gonna take us something like five years maybe to to start deploying robot. And we deployed our first robots, I think, eighteen months into the company. So it either means that I'm very bad at at making those predictions or it's been just moving much, much faster. There's been multiple of these moments. I think our first release, Pi Zero, was definitely one of those moments where I thought it would take a very long time before we have robots folding laundry, especially diverse pieces of laundry fully autonomously. This was one of these holy grail of robotics that I thought were gonna... It's gonna take many, many years before we get to. And we got there within the first seven months of the company. I think the next model was another one of these moments where I thought it was gonna take very, very long to get robots to perform in an environment that they've never seen before. And with our Pi 0.5 release, we got we got robots to new homes that they've never seen before. And home is kinda like the most diverse environment you can imagine. So it's the hardest version of that challenge. And that started working as well. Then with with PyStar six, we started seeing that these robots can perform tasks at very high clip, at very high success rates. And now we have videos of robots doing extremely dexterous tasks like making coffee on the proper espresso machines, machine for like thirteen hours without any cuts, just like thirteen hours straight. The robot is doing coffee over and over again and cleaning after itself. So I think all of this is just moving very, very fast, and every single release so far has been quite surprising to me. I thought it was gonna take longer or that the problem is much more difficult than I than I expected.
01:04:06 · Mario Gabriele
[01:04:06] S1: But I think that's maybe one thing about working with these models. I think probably people working on LLMs would have similar answers that they thought it was gonna take much longer to get to a certain level of capability. You know, when you're running a robot for thirteen hours straight making espresso, what does a good failure rate look like in that in that case, or what are you excited by when you when you see that happening, if anything? And then, you know, something I was talking about with with a friend who's in this industry is that a lot of the times, the failure for these kinds of robots is maybe less obvious and more like it almost doesn't realize it's stuck in some way and and doesn't find a way out of of being stuck in that loop. Is that is that sort of something that you you see a a fair amount in this case?
01:04:54 · Karol Hausman
[01:04:54] S0: I think in general, don't pay enough attention to this problem of liability. And that that's maybe a big difference between something like a language model or a chatbot and a robot. When you talk to a chatbot and it makes a small mistake or it just like puts not quite the right word, it's totally fine because you're right there and you can absorb all of this. You can correct it. You can ask it again, or you can you can interpret it the right way. You're very forgiving. And the physical world is is the opposite of that. That's not very forgiving. If you make any small mistake, if you know, if you grab the porta filter just like a little bit off or you don't insert it just quite right, then the whole task is going to fail in a catastrophic way. So I think the the bar on performance is just much much higher because there isn't someone to help you, someone to catch you. And I think this is a very big problem. Now the the other thing you point out is more about the model being self aware of where it's where it's not going well. And the power of reinforcement learning is that it very often needs to have an additional function that tries to predict how close you are to success. We call this a value function. And that value function should have an understanding of how well you're doing so that then you can improve upon it. And what we found so far is that you can train these value functions to be pretty good. You can train them the extent where we don't fully know that the robot is about to fail yet, that everything is looking okay, but the value function already knows. And they start seeing that, like, the the the the expectation of of of success is starting to go down. This is actually something that we've also seen with with works like AlphaGo, where we had a computer program learn how to play Go. Where if you if you watch some of the games that this AI system had against the the the champion at at Go, What was quite intriguing is that in many of these games, you have world's experts at Go talking about and commentating on the game. They're all thinking that the game is basically head to head that's like, you know, it's very unclear who is winning, whether it's AI or Lee Sedol. Then But you look at the value function predictions, and it's basically over. Like, the the AI system knows 100% that the game is over.
01:07:10 · Karol Hausman
It's already won by AI, and Lee Sedol has no chance. But the entire world, you know, all the experts, including probably Lee Sedol himself think that that this is very, very close game. So I think we will probably get to to similar models in in robots where they will be able to predict how well they're doing or how likely the success is much better than we will. One of