Digitized | Spotlight
00:16:39 · Michael Cho
[00:16:39] S2: Yeah. That's... Yeah. Well... Yeah. So the the competition basically, fundamentally, we we take the top human gamers who play our game to take on AI models that are built by all these different labs. Know, this could be universities like UT Austin, National University of Singapore, Seoul National University, basically like labs that do this kind of research. So they built AI models that basically take on the same game as a human gamer. So if you think about it, it's basically humanity versus AI playing the same basically Pokemon GO game, something like that. And so like I mentioned earlier, the researchers are very interested in this platform because this is... We have basically one of the largest... Mean, far... Actually by far the largest dataset of its kind already, and many of the research team takes take advantage of that to build the models. And then more importantly, because these robots that we've built are really affordable, cheap, therefore a lot easier for them to actually start experimenting versus some of these other traditional robot like you know they're very fancy, way more capable than ours, it's just that they're damn expensive. Yeah it's so I'm really happy to see so now there are like academic papers that have been being written. Many of them have been published and more on the way, and so we are... I feel that we are really making a meaningful difference already through the data set and also through through this competition. By the way, like, the humans are way better than the AI in industry marketing. So so so the AI is still pretty shitty.
00:18:23 · Winnie, Michael Cho
[00:18:23] S1: We're not quite there yet. Our job is still backstage from AI. [00:18:28] S2: Yeah. I mean, the urban navigation... Business sidewalk driving is in many ways the same kind of like stack as a problem set as self driving on the road. It's just that way more resources and also way more data have been gathered on the road and so cyber robot in a way is way behind. So but hopefully, through what we've been doing, are getting a lot of these researchers to push further ahead. I think if you look at the universities and also some of our co organizers, many of them actually, like you mentioned, from researchers from DeepMind, Meta AI, and and universities. So I think for them to be involved with us, I mean, speaks volume about, like, this is very much needed in the research world, and and I think we we are already providing value even at this stage. Yeah. [00:19:23] S1: Yeah. No. I think you also run a podcast with... Where you kind of talk a lot about and interview some of these top researchers, correct, in robotics. What are some, like, exciting frontier things? I guess a lot of what you're doing is already quite frontier, but, like, what are some interesting research pathways or, you know, exciting things that you guys have been discussing lately that you can share?
00:19:23 · Winnie, Michael Cho
[00:19:23] S1: Yeah. No. I think you also run a podcast with... Where you kind of talk a lot about and interview some of these top researchers, correct, in robotics. What are some, like, exciting frontier things? I guess a lot of what you're doing is already quite frontier, but, like, what are some interesting research pathways or, you know, exciting things that you guys have been discussing lately that you can share? [00:19:46] S2: Yeah. I mean... So... Yeah. Me me and this guy, Chris Paxton, shout out to Chris. He... He's a real expert. He himself is a top researcher. So I I feel very privileged to do the podcast with him. I think both of us, I mean it's, I mean you can tell me I'm very geeky, read like research papers for years and people are going to sleep so I've been doing it for ages. [00:20:07] S1: I mean you spent COVID building robots so.
00:20:10 · Michael Cho
[00:20:10] S2: Yeah yeah yeah so that was my COVID. Yeah so yeah there are I think there are some clear trends I would say because like we do this like once or twice a week now there's just so many cool papers coming out But broadly speaking, like, I I can name a few big trends. Like, for... So for example, models that are trained using ecocentric videos. This is... I I I I kinda feel that the last maybe half a year, there was a big trend. There's a lot more papers coming out partially because of some of these, like, work that's done especially by Meta. They have this smart glasses called ARIA. So that has spun off a bunch of research. And to be honest, EcoCentric video had been looked at by the computervisioncom... Community for, like, ages, But I thought nothing really moved until quite recently, and now there's a huge amount of interest. So something has changed, I think, technology that that allows researchers to pick a lot... To do a lot more with ecocentric video. So I think that's one big one. Of course, all the humanoids, they are very, very... Like, advancing very, very quickly. You know, a year plus ago, they can barely walk. Now they are, like, kicking around, like, you know, doing crazy dance, and and and so, I think there are a lot of very cool papers coming that's specific to humanoid as well. Some of the other stuff also more fundamental research like I think reinforcement learning, it's traditionally doesn't really work for robotics. Know? So, like, for example, AlphaGo is is basically all RL, but but because you you can always, you know, like, we can... You can restart the game. It's all in simulation. Right? And so whereas in the robot in the real world if you do some rl like on policy rl online rl and then basically if robots fail you can't you can't respawn yeah so that's why it's never quite worked but I'm starting to see people pushing a boundary on what you can do and I think that's a very big piece of the puzzle as well that needs to be solved. It's not soft but a lot of cool works coming out. I just saw yesterday like something from physical intelligence using IRL to a great extent.
00:22:24 · Michael Cho, Winnie
So, and by definition, I mean by extension there are these like in context learning, things like that that push, allow the robots to learn a lot more when they're being deployed instead of just the current paradigm which is just calling a huge amount of data first and then you train a model, then you deploy and then you see how it feels, then you train a huge model again, which is super inefficient. That's not how animals and humans do it, We just learn as we go, [00:22:54] S1: iterate as we go, [00:22:55] S2: and then we just accumulate. Exactly, so the robots, most of the, exactly, so I think these are cool research trends and we have a couple of many authors that cover these different things, yeah. So these are some of big trends, I would say. [00:23:10] S1: Love that. Earlier, you mentioned how humanoids are getting a bit better. I am curious because I know you guys also incubated and spun off this, like, humanoid fighting entertainment franchise. And tell me more about that. Like, I feel like there's, like, a lot of sold out shows. There's a lot of coverage on that. What led you to launch Yep. UFB, I believe it's what's called, and kind of, like, how does that tie to BitRobot?
00:23:35 · Amanda
[00:23:35] S0: So for robots incubated UFB, it's a yeah, like you said, it's a humanoid fighting league. And yeah, they've had a bunch of shows in SF and LA sold out covered in New York Times, Ashley Vans, Washington Post. I think they had... Now they're in LA working with creators like Wiz Khalifa that are piloting the bots. So, yeah, they've gotten a lot of attention. I think, like, and I think Michael can talk about like his thesis around robotic gaming. But essentially like, if you think about what is the first use case for, for robots like humanoids, when they're not super, let's say like, reliable in doing like everyday tasks, it's entertainment. And so like, I know Michael views that the first big use case and like where most of the purchases of humanoids are going to go to is towards entertainment. And so how can you use entertainment to then create a flywheel for robotics R and D? So use the media and entertainment franchise to finance the bots and then use the bots when they're not fighting for a a subnet. In this case, something like humanoid navigation is is what we're exploring. So that's the concept. I know, like, Michael's the kind of the visionary though behind UFB, so I don't know if you wanna add.