Going Direct Conversations
00:11:33 · Jordan Wolfe, Chris Paxton
[00:11:33] S2: So let's talk about those bottlenecks, right? So obviously, it's common knowledge that obviously there's a training gap. And obviously, it's like you need the most diverse, large data sets and to deal with edge cases and with larger context windows. We can go on and on and on. Can we drill, I want to get each of your perspectives on really where the main gap lies and the main bottleneck lies because it's so hard to determine between what, especially folks that are building robots put out there from a marketing standpoint and how far away we are to actually make them useful. So Chris, maybe I'll go over to you first. If you were to boil down if there's one or two things that really stand out to you of... the bottleneck around the current bottleneck around training gaps and, you know, where are we really at? What is the current state of the industry? [00:12:26] S0: I mean, I think that the hardest things to do are to get robots out there into the real world and make them do stuff, which, I mean, it's kind of circular, right? But, like, getting robots and deploying them and making them work reliably and robustly and survive. It's like, can you make the robot actually work for a long period of time? It's like, I mean... We need data, but that's secondary to getting robots out there. I think a lot of the time, making them work. Once they work, the data comes in naturally. They're always producing it. [00:13:05] S2: Michael, what's your perspective?
00:13:09 · Michael Cho
[00:13:09] S1: Yeah, I think broadly speaking, I see two huge bottlenecks in robotics. One is obviously lack of robotics data. I think that, so if you look at the internet, like I think that the whole internet is just the output of humans' conscious thoughts, right? So we're able to record those things down. That's why those are great training models for LLM, VLMs and whatnot. But the way humans like, you know, move in a physical world, like the way I move my water bottle, I can only show you. but I can't record it at, let's say, 30 hertz, you know, the exact angle, the velocity and all that. So these action labels, because we can't record it, so that's why the data is totally missing. And except, I guess, maybe Tesla, they put a human in a robot on wheels and then they call it what the human driver is doing. So those action labels. So I think the robotic data is very much needed and there are three types of robotics data, like tablet operation and video data and also like synthetic or simulation data. So that's one big bottleneck, just data and three types of data. The other huge bottleneck I see is real-world evaluation. And I think this is something that is very unique to physical AI. In the digital AI run, like if you, let's say, if DeepMind or OpenAI, if they drop another reasoning model, typically, you know, obviously they have a bunch of benchmarks. And if you look at all those benchmarks, right, like for example, they will take, let the AI do, let's say, a math Olympian test. That test can be done like just like that, right? So even though there's no perfect benchmarks, getting the benchmarks is never a bottleneck in the iteration cycle. Whereas the physical AI, like you can test all you want in simulation, but ultimately, like literally rubber hits the road, right? You know, you need to see how they fail. And because you can't speed up time, Like, how do you know, for example, whether you build a grid model unless you measure how long does it take before it fails? Like, this is the only way to do it. And so I think real world eval would be a huge bottleneck. Literally, if you look at robotics, I mean, Chris and I talk about this all the time, there's no benchmarks. You know, there's no benchmarking because it is just nearly impossible to have benchmarks in robotics. Whereas everybody's robots difference.
00:15:25 · Chris Paxton, Michael Cho, Jordan Wolfe
[00:15:25] S0: Everybody's house is different. [00:15:27] S1: Yeah, exactly. In that paper, they have this success rate. You just have to take it that, okay, they did their work and it is what they say, but no one can verify it. [00:15:38] S0: Sure. It's getting better, but yeah, definitely not. Oh, that's so hard. Yeah. [00:15:42] S2: All right. Well, it's an interesting question, right? I guess this one's for you, Chris, because obviously you're working on or agility is working on the humanoid form factor. And it's kind of like when the world of robotics is like, just show me. It's very different than some abstract or benchmarking process to benchmark LLMs. [00:16:07] S0: Is that okay? Yeah. The thing is that you can't just get a good feel for how a robot works from a video because you can always get a good video. Honestly, the other thing is that I think people have seen lots of really good robot videos for a long time because they've seen Star Wars or whatever. You've seen robots that can do things that they can't even do now. I think you have to be able to show robots in the real world and see it for yourself to be able to see what it can do. I think a lot of videos can be misleading. I wanted to drill into the different types of training, methods for training.
00:16:07 · Chris Paxton, Jordan Wolfe
[00:16:07] S0: Is that okay? Yeah. The thing is that you can't just get a good feel for how a robot works from a video because you can always get a good video. Honestly, the other thing is that I think people have seen lots of really good robot videos for a long time because they've seen Star Wars or whatever. You've seen robots that can do things that they can't even do now. I think you have to be able to show robots in the real world and see it for yourself to be able to see what it can do. I think a lot of videos can be misleading. I wanted to drill into the different types of training, methods for training. [00:16:44] S2: I wanted to get your perspective on it. Because that's also another aspect that's kind of a little bit unclear. What is the most effective? Is it a mix of all of these? Let's start with just teleoperation. Michael, what's your perspective on what does teleoperation truly allow you to do to solve some of the training gaps that currently exist?
00:16:44 · Jordan Wolfe, Michael Cho
[00:16:44] S2: I wanted to get your perspective on it. Because that's also another aspect that's kind of a little bit unclear. What is the most effective? Is it a mix of all of these? Let's start with just teleoperation. Michael, what's your perspective on what does teleoperation truly allow you to do to solve some of the training gaps that currently exist? [00:17:08] S1: Yeah, I would say teleoperation is the gold standard amongst the three types of data I mentioned just now. But obviously, it's extremely costly to collect. I say it's a gold standard because you are doing in the real physical world with a real physical robot. So you don't suffer from, let's say, embodiment gap. And you collect exactly, typically, I mean, this is humans controlling the robot, right? So you also, in a way, collect human intention, like what would a human do, like in this instance, right? So there's some social compliance stuff that also naturally get collected. So I think it's great for all those reasons. But of course, these robots break all the time, you know, and you can't respawn them, you know, like in a simulation. So that's a huge thing. Yeah. So I, I think my thing is that if you can gamify it, then, then hopefully you get more people to constantly do it, but you still cannot get away from the fact that these things will break all the time and makes it very, very hard to scale. [00:18:08] S2: Yeah. Yeah. What's your thought, Chris, on the teleoperation side?
00:18:08 · Jordan Wolfe, Chris Paxton
[00:18:08] S2: Yeah. Yeah. What's your thought, Chris, on the teleoperation side? [00:18:13] S0: I mean, I think that, yeah, like you're saying, that's the gold standard, getting real robot data. And I think the gold standard is actually probably, the best data is probably the data the robot collects itself. It's probably like reinforcement learning kind of data. But that's pretty rare still. But the robots are... Teleoperation data is great as long as you can actually get it, right? But while scaling, it's hard. And so that was actually the next question was kind of reinforcement learning. [00:18:45] S2: So Chris, can you kind of explain that for folks that might not be so understanding of what that actually means in the term of robotics space?