Meta's Mimic: Teaching AI Agents to Move Like Toddlers

0:00The holidays are here and that means it's the most wonderful time of the year to save with Rakuten. Use Rakuten to stack cashback at your favorite stores on top of holiday sales. That's savings on savings. With Racketing, you get cash back on gifts for everyone on your list. From toys for the kids, to kitchen gear for the person who loves to cook, to electronics for everyone. You can even save on something for yourself. Just shop the stores you love and cash back is automatically added to your account. And you can get paid with gift cards, PayPal or check. or eligible American Express card members can even choose to earn membership rewards points instead of cashback.

0:41It's truly a no-brainer. Join for free today and get a new member bonus after minimum qualifying purchases. Just go to Rakuten.com, download the app, or install the browser extension. That's R-A-K-U-T-E-N. Terms and conditions apply. What can 160 years of experience teach you about the future? When it comes to protecting what matters, Pacific Life provides life insurance, retirement income, and employee benefits for people and businesses building a more confident tomorrow. Strategies rooted in strength and backed by experience. Ask a financial professional how Pacific Life can help you today. Pacific Life Insurance Company, Omaha, Nebraska, and in New York.

1:24Pacific Life and Annuity, Phoenix, Arizona. One of the big breakthroughs that I think AI is going to make in the very near future is moving beyond just being something like OpenAI's ChatGPT and actually having that intelligence put into robots, right? I think right now a lot of people are concerned about how ChatGPT is going to replace jobs in the white collar space, right? It's going to replace marketers writing articles and all sorts of other things we've discussed on the podcast. I believe the next phase after this goes to actually robots that are going to start replacing blue-collar workers.

1:58I'm not saying this is good. I'm not saying this is bad. I'm just saying what I believe is going to happen, right? So inevitably, we're going to have humanoid robots that are trained specifically to be an engineer or to be an electrician or to be a plumber. All the jobs we thought, hey, at least AI is not going to replace these. I believe it's going to go there. But the big problem that these humanoid robots are currently facing is, well, being humanoid robots. Walking around is very difficult. While Chai Chibutti may have created, you know, a brain that's going to get more and more powerful over time, we still don't have the host to put that brain in, the actual robot itself.

2:33So today on the podcast, we're going to be talking about how Meta's new AI agent is learning to move by copying toddlers. And this is a biomedical model that learns like humans. and this is essentially designed to help robots and humanoid robots learn how to walk. So today on the podcast, we're going to be talking about what they're training here, what the implications are, and why this is important. So in a bit of a state-of-the-art simulation, a skeleton arm, powered by artificial intelligence, expertly handled a toy elephant, rotating it in its grip. And essentially, this was mimicking the exploratory actions of a curious toddler.

3:14It used 39 muscles across 29 joints, and this innovative system then proceeded to interact with various objects, a toothpaste tube, a stapler, an alarm clock. Similarly, in another scenario, skeletal legs propelled by 80 muscles through 16 joints engaged in motor actions reminiscent of a toddler's first steps. So all of these feats were achieved by the Myosuite platform, which is particularly the latest edition of this software, so Myosuite 2.0. And essentially this unveiling came from a collaboration between Meta-AI and prominent researchers at McGill University in Canada, Northeastern University USA, and the University of Twent, Netherlands.

4:00So by focusing on machine learning for biomechanical control, the aim is to emulate human level dexterity and agility. So I think that the complexity involved in coordinating large and small muscle groups poses significant control challenges here. But I believe that Myosuite offers a really valuable repository of musculoskeletal models and benchmarks tasked for open research. So Mark Zuckerberg actually highlighted the platform's potential saying, quote, this research could also help us develop more realistic avatars for the metaverse. I think this is, I don't know people's opinion on that. In my opinion, it's like, oh my gosh, we're recreating like skeletons and how like muscles move and stuff.

4:47And Zuckerberg's like more realistic, like avatars in the metaverse. I'm like, dude, we literally could like recreate humanoid robots with like every muscle and tendon that like actually function and move exactly how we move. or like someone that has their actual arm amputated or like a leg amputee could get like an actual arm or muscle that functions like 100 % like an actual one. Like this is incredible for science and research and Zuckerberg's all caught off on like the metaverse, whatever. It's just dumb. Anyways, Vakash Kumar, who's a prominent researcher on the project, he drew a couple comparisons between the intricate mechanics of the human body and traditional robots.

5:26So, while robots typically have a single motor per joint, the human anatomy involves multiple muscles per joint, with these muscles interacting through various joints. So, the challenge isn't merely about, you know, initiating motion, but really substantiating intricate patterns of muscle activation. This is a feat that the human brain accomplishes effortlessly, and so for Kumar, attempting to replicate these strategies in Myosuite really exceeds the simplicity of, you know, directing robots. He is very optimistic about it, though. He's essentially drawing from human biomechanics, and he thinks that this could provide pivotal insights.

6:09So Kumar believes that evolution shows our complex anatomy for profound reasons, emphasizing, quote, if an easier solution was possible, it would be foolish for evolution to converge on this complicated form factor. So Kumar has recently transitioned to a full-time role at CMU's Robotics Institute from a dual role as a meta researcher and adjunct professor. But I think really the inception of this innovation is credited to Meta AI's fundamental AI research, which is FAIR division. And so when MayoSuite 1.0 was launched back in May of 2022, Zuckerberg himself hinted at its potential implications for Meta's product line.

6:50and I mean okay I guess I'll say one other thing on this right it was like we're being like it's gonna make avatars of the metaverse more realistic whatever like if that's what it takes for Zuckerberg to fund this right and we see like a lot of incredible breakthroughs in science and it helps people in a lot of ways that like whatever if it took more realistic avatars in the metaverse and I mean at the end of the day they're funding this amazing research that's gonna do some cool stuff whatever I'll I guess I'll take it right so I think that in 2022 a contest named Mayo Challenge was organized and it was concluded at the Neural IPS AI conference and teams essentially grappled with stimulating a hand to rotate dice and manipulate some boating balls.

7:34So though many displayed remarkable proficiency, Kumar noticed limitations to their generalization capabilities. So a change in the object's properties posed significant challenges to the AI algorithm, right? Like if it's going to be a ball or a dice or a banana, it's very difficult for these robots to switch between objects and treat them in, you know, impressive ways. So recognizing this, the meta team really kind of aimed to design AI agents that were proficient in generalizing across tasks, which is employing, you know, the Mayo arm and Mayo legs as training grounds. And so Kumar's strategy was to train the algorithm with diverse object representations.

8:12And I think doing this really enhanced its task-specific learning rate. So as detailed in a paper for the International Conference in Machine Learning, this method, similar to a toddler's exploration, proved to be really effective, which is really exciting that they are making some breakthroughs here, and that these new algorithms they're training in AI are actually helping it to learn how to switch between a lot of different objects, and it was quite effective in that. So a recent publication at the Robotics Science and Systems meeting detailed similar success with myolegs. Vitro Cagani, who's a researcher at Meta, remarked on the broader implications suggesting that the core knowledge is transferable across systems.

8:53That's really, really exciting, right? The core knowledge that, like, essentially the algorithm they use to make the hand with all of its muscles and joints be able to pick up different objects transferred across systems, meaning it was able to go into the leg models that they had and teach the legs how to walk across different situations. So this is really, really exciting to know. I think this year's Mayo Challenge 2023 is going to be concluded at Neural IPS, and it has set advanced tasks. So participants have to utilize the Mayo arm for handling household objects and the Mayo legs in a game of tag.

9:27So Imo Tordove from the University of Washington commented on Mayo Suites, emphasis on broader kind of representations noting its potential utility across a lot of different tasks so he compared their approach to neurosciences muscle synergies principles which essentially emphasizes group muscle activations for efficiency so Torodo really commended Myosuite's ability to build these kind of models from scratch but I do think that the meta team as they're moving forward very rapidly I think they're going to need to refine the AI models a lot further I think emulating toddler toddlers requires more than just kind of handling and mobility i think true comprehension might only come when these ai models can actually taste their surroundings much like a toddler's you know instinct to mouth objects or whatever right like right now these are all computer softwares and i think they're going to um need to bring these into the real world to train them but i do think that we're going to see some really incredible um results as this happens and this has big implications not just for avatars and the metaverse but I think for science for artificial limbs and for all sorts of things that are incredibly exciting at the moment