[
  {
    "startMs": 1309,
    "text": "[music]"
  },
  {
    "startMs": 12559,
    "text": "I think uh all of you know here that how"
  },
  {
    "startMs": 15040,
    "text": "much progress has AI made in the last"
  },
  {
    "startMs": 16960,
    "text": "five years. Uh seems like it's more than"
  },
  {
    "startMs": 19920,
    "text": "uh last 100 years of progress. And what"
  },
  {
    "startMs": 23119,
    "text": "is happening especially in AI is AI came"
  },
  {
    "startMs": 25760,
    "text": "for language then speech audio video and"
  },
  {
    "startMs": 28960,
    "text": "everybody's excitement is next is"
  },
  {
    "startMs": 30880,
    "text": "robotics okay and this excitement has is"
  },
  {
    "startMs": 33840,
    "text": "going on peak uh this year uh for some"
  },
  {
    "startMs": 37280,
    "text": "reason which is still beyond my uh"
  },
  {
    "startMs": 39680,
    "text": "understanding and you can also see uh"
  },
  {
    "startMs": 42399,
    "text": "leaders like Jensen talking about"
  },
  {
    "startMs": 44160,
    "text": "physical AI as the next frontier of"
  },
  {
    "startMs": 45600,
    "text": "robotics. So it seems like if you open"
  },
  {
    "startMs": 48079,
    "text": "Twitter or LinkedIn, it seems like"
  },
  {
    "startMs": 49760,
    "text": "robots robotics is already here and"
  },
  {
    "startMs": 51920,
    "text": "we'll have home robots in our homes"
  },
  {
    "startMs": 53600,
    "text": "sometime by December uh as uh as it's"
  },
  {
    "startMs": 56879,
    "text": "being said by multiple people. But I'm"
  },
  {
    "startMs": 59120,
    "text": "here to highlight that robotics has been"
  },
  {
    "startMs": 62399,
    "text": "almost here for the last 70 years"
  },
  {
    "startMs": 66155,
    "text": "[sighs]"
  },
  {
    "startMs": 66799,
    "text": "now. Those of you who got into robotics"
  },
  {
    "startMs": 69600,
    "text": "in the last four or five years, go and"
  },
  {
    "startMs": 72159,
    "text": "take any any any course in robotics and"
  },
  {
    "startMs": 75280,
    "text": "if you can distinguish the videos in"
  },
  {
    "startMs": 77439,
    "text": "robotics from 30 years ago versus today,"
  },
  {
    "startMs": 80640,
    "text": "uh I owe you a dinner. So to give you an"
  },
  {
    "startMs": 84320,
    "text": "example, let's look at this robot. So"
  },
  {
    "startMs": 86240,
    "text": "this robot here is an image and the goal"
  },
  {
    "startMs": 88159,
    "text": "of this robot is to look at this image"
  },
  {
    "startMs": 90479,
    "text": "of blocks and arrange these blocks in"
  },
  {
    "startMs": 93680,
    "text": "front of it so that it looks the same"
  },
  {
    "startMs": 96240,
    "text": "pattern as in the image. Okay, this is"
  },
  {
    "startMs": 98720,
    "text": "an extremely simple task. But if you"
  },
  {
    "startMs": 100479,
    "text": "think about this, it's a 3D block. You"
  },
  {
    "startMs": 102079,
    "text": "have to look at understand 3D from each"
  },
  {
    "startMs": 103520,
    "text": "side. And the robot can do this pretty"
  },
  {
    "startMs": 106159,
    "text": "precisely. Any guesses how old this"
  },
  {
    "startMs": 108799,
    "text": "result is?"
  },
  {
    "startMs": 111280,
    "text": "Anyone? It's a guess. You can make any"
  },
  {
    "startMs": 112799,
    "text": "guess. Huh?"
  },
  {
    "startMs": 114560,
    "text": ">> 40 years. Okay, that's the highest."
  },
  {
    "startMs": 117901,
    "text": "[panting]"
  },
  {
    "startMs": 118960,
    "text": "This is from 1960s. This is way before"
  },
  {
    "startMs": 122079,
    "text": "there were computers. Uh, okay. So, this"
  },
  {
    "startMs": 124799,
    "text": "is this is called MIT copy demo. Uh,"
  },
  {
    "startMs": 126880,
    "text": "this was the start of uh this is this"
  },
  {
    "startMs": 129360,
    "text": "predates AI. Uh, as you know, today look"
  },
  {
    "startMs": 132720,
    "text": "at this one."
  },
  {
    "startMs": 137599,
    "text": ">> One exhibit at the Nuclear Congress in"
  },
  {
    "startMs": 139360,
    "text": "Philadelphia has the perfect formula for"
  },
  {
    "startMs": 142000,
    "text": ">> what's happening here. The guy behind"
  },
  {
    "startMs": 143920,
    "text": "the scene is controlling these robots, a"
  },
  {
    "startMs": 146319,
    "text": "leader follower system. And if you look"
  },
  {
    "startMs": 148480,
    "text": "at any big lab, any big company, any big"
  },
  {
    "startMs": 151599,
    "text": "academic lab, how they get data, they"
  },
  {
    "startMs": 154000,
    "text": "use teley operation system, which"
  },
  {
    "startMs": 155760,
    "text": "follows the exact same principles."
  },
  {
    "startMs": 158480,
    "text": "Any guesses for the year for this one?"
  },
  {
    "startMs": 162480,
    "text": ">> Now, everybody would correct"
  },
  {
    "startMs": 164000,
    "text": "aggressively on the year, but this is"
  },
  {
    "startMs": 166400,
    "text": "from 1957, 68 years ago. Okay? Ask"
  },
  {
    "startMs": 171440,
    "text": "anybody in your family who is older than"
  },
  {
    "startMs": 173440,
    "text": "actually they may not be around like so"
  },
  {
    "startMs": 176080,
    "text": "because this is for 60 you have to ask"
  },
  {
    "startMs": 178319,
    "text": "somebody who is 80 year old what was the"
  },
  {
    "startMs": 180239,
    "text": "state of technology at that time like"
  },
  {
    "startMs": 182879,
    "text": "RAM for few KB of RAM you would have a"
  },
  {
    "startMs": 185280,
    "text": "size of a room uh like this big of a"
  },
  {
    "startMs": 187519,
    "text": "setup and this is from that time when"
  },
  {
    "startMs": 189440,
    "text": "people could make it work on not just"
  },
  {
    "startMs": 192080,
    "text": "this you keep coming back every decade"
  },
  {
    "startMs": 194400,
    "text": "the videos look as cool like here is"
  },
  {
    "startMs": 196319,
    "text": "robot a humanoid juggling playing"
  },
  {
    "startMs": 199040,
    "text": "foosball uh with human and this is also"
  },
  {
    "startMs": 203040,
    "text": "from you can see this is all they were"
  },
  {
    "startMs": 204640,
    "text": "all predate deep learning this from 8"
  },
  {
    "startMs": 206560,
    "text": "years ago so this robot is a result from"
  },
  {
    "startMs": 208480,
    "text": "Berkeley can clean up the whole table"
  },
  {
    "startMs": 210480,
    "text": "can arrange items in a in a in a box"
  },
  {
    "startMs": 213120,
    "text": "this was done by one grad student with a"
  },
  {
    "startMs": 215200,
    "text": "single GPU machine nothing more than"
  },
  {
    "startMs": 217120,
    "text": "that okay now this picture of and if I"
  },
  {
    "startMs": 220319,
    "text": "colorize all these videos from past and"
  },
  {
    "startMs": 222959,
    "text": "apply genai filter for the modern voice"
  },
  {
    "startMs": 225680,
    "text": "you cannot tell that this [laughter]"
  },
  {
    "startMs": 227519,
    "text": "video is from 19657 or this is from"
  },
  {
    "startMs": 229760,
    "text": "today And these are not even the oldest"
  },
  {
    "startMs": 231680,
    "text": "results the oldest ones go to 1940s."
  },
  {
    "startMs": 234319,
    "text": "Okay. So what is this that if you look"
  },
  {
    "startMs": 238560,
    "text": "at last 70 years tech every other"
  },
  {
    "startMs": 240879,
    "text": "technology every other technology has"
  },
  {
    "startMs": 243200,
    "text": "come way far language understanding"
  },
  {
    "startMs": 245599,
    "text": "computer vision mobile phone chips why"
  },
  {
    "startMs": 248159,
    "text": "is robotics is stuck in this primitive"
  },
  {
    "startMs": 250159,
    "text": "age uh for for this long and the reason"
  },
  {
    "startMs": 252319,
    "text": "is uh a general brain. Robotics has"
  },
  {
    "startMs": 255680,
    "text": "always been approached as a hardware"
  },
  {
    "startMs": 257359,
    "text": "problem from ground up and this the"
  },
  {
    "startMs": 259759,
    "text": "reason that everything around robotics"
  },
  {
    "startMs": 261919,
    "text": "has progressed but robotics is still"
  },
  {
    "startMs": 263440,
    "text": "stuck in the same land from the last 70"
  },
  {
    "startMs": 265120,
    "text": "years. So there's a very famous paradox"
  },
  {
    "startMs": 266960,
    "text": "called Morave paradox. I don't know if"
  },
  {
    "startMs": 268720,
    "text": "you know about Moravec. He was also a"
  },
  {
    "startMs": 270080,
    "text": "CMU professor and I'm also CMU professor"
  },
  {
    "startMs": 272720,
    "text": "so have to quote him for sure. He was"
  },
  {
    "startMs": 274639,
    "text": "one of the founding figures in in AI."
  },
  {
    "startMs": 276880,
    "text": "After 30 years of working in robotics he"
  },
  {
    "startMs": 279280,
    "text": "arrived at a very simple conclusion."
  },
  {
    "startMs": 281360,
    "text": "Hard is easy, easy is hard. Okay."
  },
  {
    "startMs": 284080,
    "text": "Whatever human believe to be hard is"
  },
  {
    "startMs": 286160,
    "text": "extremely easy for computers and vice"
  },
  {
    "startMs": 288400,
    "text": "versa. And you can see it happening"
  },
  {
    "startMs": 290160,
    "text": "right in front of your eyes. You would"
  },
  {
    "startMs": 291919,
    "text": "say doing math is hard. Math Olympiad"
  },
  {
    "startMs": 294639,
    "text": "gold medal is really really hard."
  },
  {
    "startMs": 296560,
    "text": "Climbing a stair oh super easy. Now look"
  },
  {
    "startMs": 299600,
    "text": "around you in technology where we are in"
  },
  {
    "startMs": 301919,
    "text": "terms of what has been solved and what's"
  },
  {
    "startMs": 303280,
    "text": "not being solved. So this is robotics"
  },
  {
    "startMs": 305040,
    "text": "for you. Okay. It's very it is not yet"
  },
  {
    "startMs": 307680,
    "text": "another application of AI. It is what AI"
  },
  {
    "startMs": 310560,
    "text": "was founded for in the very beginning"
  },
  {
    "startMs": 312160,
    "text": "and has made very little progress"
  },
  {
    "startMs": 313440,
    "text": "towards. This is not yet another"
  },
  {
    "startMs": 315360,
    "text": "application where deep networks can come"
  },
  {
    "startMs": 316960,
    "text": "in and attack the field. This is has to"
  },
  {
    "startMs": 319360,
    "text": "be thought of with fundamentally first"
  },
  {
    "startMs": 321440,
    "text": "principles from the ground up. This is a"
  },
  {
    "startMs": 323919,
    "text": "very famous example of where you can"
  },
  {
    "startMs": 325840,
    "text": "have a you know a computer beat uh Gary"
  },
  {
    "startMs": 328880,
    "text": "Casper on chess in '90s but you can you"
  },
  {
    "startMs": 331360,
    "text": "still don't have a a a computer or robot"
  },
  {
    "startMs": 333919,
    "text": "that can pick up the chess pieces and"
  },
  {
    "startMs": 335600,
    "text": "arrange them on any chess board. Now,"
  },
  {
    "startMs": 340000,
    "text": "so how do we go about solving this?"
  },
  {
    "startMs": 341840,
    "text": "Okay,"
  },
  {
    "startMs": 343520,
    "text": "so on a more positive note, uh we have a"
  },
  {
    "startMs": 347120,
    "text": "magic sauce for AI success, right? You"
  },
  {
    "startMs": 349680,
    "text": "get big data set, you train big models"
  },
  {
    "startMs": 352320,
    "text": "and magic happens. Okay, now we know"
  },
  {
    "startMs": 354639,
    "text": "this template is working well very well"
  },
  {
    "startMs": 357759,
    "text": "in many topics. Okay, now can we apply"
  },
  {
    "startMs": 361360,
    "text": "the same recipe to robotics? Well, it's"
  },
  {
    "startMs": 363120,
    "text": "not uh directly applicable because we"
  },
  {
    "startMs": 366240,
    "text": "have no data. There is no internet of"
  },
  {
    "startMs": 368160,
    "text": "robotics data and as I said earlier you"
  },
  {
    "startMs": 370639,
    "text": "can go and collect data uh manually on"
  },
  {
    "startMs": 372720,
    "text": "the robot called telly operation and as"
  },
  {
    "startMs": 375520,
    "text": "you notice teley operation is not 5 year"
  },
  {
    "startMs": 377520,
    "text": "old thing this is a 68 or 70 year old"
  },
  {
    "startMs": 380800,
    "text": "thing so the the funny part is if you go"
  },
  {
    "startMs": 384479,
    "text": "back and you look at the whole evolution"
  },
  {
    "startMs": 385919,
    "text": "of GPD3 GPD4 models go back to GPD3 uh"
  },
  {
    "startMs": 390240,
    "text": "three years ago seems like a like a"
  },
  {
    "startMs": 392160,
    "text": "lifetime ago GPD3 only began working and"
  },
  {
    "startMs": 395360,
    "text": "caught people's attention when you could"
  },
  {
    "startMs": 397120,
    "text": "train those models on trillions of"
  },
  {
    "startMs": 399759,
    "text": "tokens. So GPD3 was already trained 30"
  },
  {
    "startMs": 401840,
    "text": "trillion plus token and today hundreds"
  },
  {
    "startMs": 403840,
    "text": "of trillions of tokens and if you"
  },
  {
    "startMs": 405360,
    "text": "collect data by manually by tell"
  },
  {
    "startMs": 406720,
    "text": "operation it takes you about 1 minute to"
  },
  {
    "startMs": 408639,
    "text": "get one example. You can do the math if"
  },
  {
    "startMs": 411199,
    "text": "you hire all of US population it will"
  },
  {
    "startMs": 413280,
    "text": "take you more than a century to reach"
  },
  {
    "startMs": 414720,
    "text": "the same scale as of GPD3 uh which is"
  },
  {
    "startMs": 417919,
    "text": "like uh and nobody uses GP3 today. Okay."
  },
  {
    "startMs": 421039,
    "text": "So this is very slow and expensive. So"
  },
  {
    "startMs": 422880,
    "text": "in robotics I would argue nobody has"
  },
  {
    "startMs": 426000,
    "text": "really scaled robotics yet and we are"
  },
  {
    "startMs": 428160,
    "text": "very far from talking about scale in"
  },
  {
    "startMs": 430080,
    "text": "robotics with the way other uh other"
  },
  {
    "startMs": 432400,
    "text": "areas have seen scale. Okay so this is"
  },
  {
    "startMs": 435120,
    "text": "where uh we have been focusing on uh our"
  },
  {
    "startMs": 437520,
    "text": "scaled is about uh 3 years old but uh I"
  },
  {
    "startMs": 440400,
    "text": "have been working on the problem for"
  },
  {
    "startMs": 441599,
    "text": "more than a decade uh that's all I've"
  },
  {
    "startMs": 443440,
    "text": "done in my career uh nothing else. Uh so"
  },
  {
    "startMs": 446160,
    "text": "at scale what our thesis is is to build"
  },
  {
    "startMs": 448560,
    "text": "what we call an omniodied intelligence"
  },
  {
    "startMs": 452639,
    "text": "any robot any task one brain okay any"
  },
  {
    "startMs": 456400,
    "text": "robot it can be a humanoid it can be a"
  },
  {
    "startMs": 458319,
    "text": "quadriped it can be a robotic arm on a"
  },
  {
    "startMs": 460000,
    "text": "conveyor belt or a dextrous hand doesn't"
  },
  {
    "startMs": 461520,
    "text": "really matter this even this hypothesis"
  },
  {
    "startMs": 464639,
    "text": "is way more general than uh one would"
  },
  {
    "startMs": 467039,
    "text": "argue humans are because we control our"
  },
  {
    "startMs": 469199,
    "text": "own body and why do we have to go so"
  },
  {
    "startMs": 471360,
    "text": "general well the argument is in robotics"
  },
  {
    "startMs": 474800,
    "text": "there there is no data anyway. So I"
  },
  {
    "startMs": 476400,
    "text": "can't pick and choose which hardware do"
  },
  {
    "startMs": 478639,
    "text": "I use data from. And we should be able"
  },
  {
    "startMs": 480800,
    "text": "to use data from any kind of hardware,"
  },
  {
    "startMs": 482720,
    "text": "any kind of task, any kind of scenario."
  },
  {
    "startMs": 484720,
    "text": "And that's the only way to truly achieve"
  },
  {
    "startMs": 486400,
    "text": "the scale of uh of what language models"
  },
  {
    "startMs": 488879,
    "text": "achieved uh three years ago. Okay. So"
  },
  {
    "startMs": 490960,
    "text": "the goal here is this uh this whole idea"
  },
  {
    "startMs": 493199,
    "text": "of any robot any task one brain. And"
  },
  {
    "startMs": 495360,
    "text": "through this talk I'll hopefully"
  },
  {
    "startMs": 496800,
    "text": "convince you why this is the this is the"
  },
  {
    "startMs": 498479,
    "text": "way to go towards robotics. Okay. So"
  },
  {
    "startMs": 500960,
    "text": "this is the uh rough intro but before I"
  },
  {
    "startMs": 503680,
    "text": "go into the more details let me show you"
  },
  {
    "startMs": 505360,
    "text": "just a teaser result."
  },
  {
    "startMs": 508800,
    "text": "So in this result, every single robot uh"
  },
  {
    "startMs": 511280,
    "text": "it's from a different company, different"
  },
  {
    "startMs": 512959,
    "text": "hardware and they're all controlled by"
  },
  {
    "startMs": 515519,
    "text": "uh skilled brain whether it's humanoids"
  },
  {
    "startMs": 518560,
    "text": "going up and down stairs, any kind of"
  },
  {
    "startMs": 520240,
    "text": "scenarios."
  },
  {
    "startMs": 522240,
    "text": "And these are not new results. They're"
  },
  {
    "startMs": 524000,
    "text": "like a couple of year old uh results in"
  },
  {
    "startMs": 526640,
    "text": "here. Robust to dust disturbances."
  },
  {
    "startMs": 531200,
    "text": "You can put them zero shot in new"
  },
  {
    "startMs": 532720,
    "text": "scenarios."
  },
  {
    "startMs": 542959,
    "text": "So the idea here any every single robot"
  },
  {
    "startMs": 546080,
    "text": "in this video takes any action anywhere"
  },
  {
    "startMs": 549600,
    "text": "in the world the brain behind the scene"
  },
  {
    "startMs": 552180,
    "text": "[music] improves because it's an"
  },
  {
    "startMs": 554240,
    "text": "omni-body brain."
  },
  {
    "startMs": 555360,
    "text": ">> Yeah, this is really hard right because"
  },
  {
    "startMs": 556880,
    "text": "like eggs are okay."
  },
  {
    "startMs": 572080,
    "text": "So that's a teaser of what we what we"
  },
  {
    "startMs": 574080,
    "text": "work on. So I want to make this talk"
  },
  {
    "startMs": 575680,
    "text": "more scientific and more uh more"
  },
  {
    "startMs": 577440,
    "text": "informative than a company ad. So I'll"
  },
  {
    "startMs": 580160,
    "text": "talk about uh how do we scale data in"
  },
  {
    "startMs": 582000,
    "text": "robotics. Okay, let's take a tour back"
  },
  {
    "startMs": 584320,
    "text": "as to how have people addressed this. uh"
  },
  {
    "startMs": 586640,
    "text": "I'll take a look back at my own career"
  },
  {
    "startMs": 588800,
    "text": "and my hypothesis for data has been"
  },
  {
    "startMs": 591040,
    "text": "changing over the years. Okay. When I"
  },
  {
    "startMs": 592880,
    "text": "began uh uh working scaling things the"
  },
  {
    "startMs": 595760,
    "text": "idea was you can have robots you can"
  },
  {
    "startMs": 598000,
    "text": "collect data for robots manually but it"
  },
  {
    "startMs": 600320,
    "text": "is too slow and robots should be allowed"
  },
  {
    "startMs": 602480,
    "text": "to collect data by themselves. So we had"
  },
  {
    "startMs": 604080,
    "text": "this whole idea of curiositydriven"
  },
  {
    "startMs": 605680,
    "text": "exploration. Uh allow robots to explore"
  },
  {
    "startMs": 608320,
    "text": "in a curious way in the environment and"
  },
  {
    "startMs": 610560,
    "text": "it has a lot of good parts like no human"
  },
  {
    "startMs": 613040,
    "text": "required. Robots can go and keep playing"
  },
  {
    "startMs": 615040,
    "text": "collect more data. We called it play"
  },
  {
    "startMs": 616560,
    "text": "data at the time. It's very rich because"
  },
  {
    "startMs": 618880,
    "text": "it has force sensors and joint angles."
  },
  {
    "startMs": 621519,
    "text": "But the difficulty is it is only in the"
  },
  {
    "startMs": 623760,
    "text": "physical world. So it's very difficult"
  },
  {
    "startMs": 625760,
    "text": "to scale. E Google was at it at the"
  },
  {
    "startMs": 628560,
    "text": "time. Uh h having hundreds of robots."
  },
  {
    "startMs": 631440,
    "text": "Even for Google it's too expensive and"
  },
  {
    "startMs": 633200,
    "text": "too slow to scale. The other idea is"
  },
  {
    "startMs": 635600,
    "text": "again telly operation. As I said it's an"
  },
  {
    "startMs": 637200,
    "text": "old idea but again the same issues"
  },
  {
    "startMs": 639839,
    "text": "impossible to scale because now you"
  },
  {
    "startMs": 641440,
    "text": "don't need robot you also need humans."
  },
  {
    "startMs": 643200,
    "text": "Uh so it's even more expensive and"
  },
  {
    "startMs": 645279,
    "text": "diversity is very limited because even"
  },
  {
    "startMs": 647920,
    "text": "if you put a robot in one setup with a"
  },
  {
    "startMs": 649680,
    "text": "human you can get thousand examples but"
  },
  {
    "startMs": 652000,
    "text": "they'll all be in the same setup. So"
  },
  {
    "startMs": 654000,
    "text": "what you ideally want carry the robot to"
  },
  {
    "startMs": 656640,
    "text": "new home every day in new scenario and"
  },
  {
    "startMs": 658560,
    "text": "that's just impossibly hard to scale."
  },
  {
    "startMs": 661680,
    "text": "Then we had major breakthrough in"
  },
  {
    "startMs": 663279,
    "text": "learning from simulation. This was one"
  },
  {
    "startMs": 665360,
    "text": "of the first result where uh one could"
  },
  {
    "startMs": 668480,
    "text": "show deep reinforcement learning uh"
  },
  {
    "startMs": 670959,
    "text": "trained in simulation transfer to real"
  },
  {
    "startMs": 672800,
    "text": "robot. This is one of one of the"
  },
  {
    "startMs": 674480,
    "text": "award-winning paper uh at the time and"
  },
  {
    "startMs": 677040,
    "text": "now it's used in every humanoid uh every"
  },
  {
    "startMs": 679600,
    "text": "company out there. uh this was from our"
  },
  {
    "startMs": 681440,
    "text": "lab at Berkeley and CMU. But again they"
  },
  {
    "startMs": 683920,
    "text": "there are pros and cons. It's very easy"
  },
  {
    "startMs": 686160,
    "text": "to scale but diversity is hard to get"
  },
  {
    "startMs": 688320,
    "text": "because you have to engineer every scene"
  },
  {
    "startMs": 689760,
    "text": "in simulation manually. So if you're"
  },
  {
    "startMs": 693200,
    "text": "listening to this like there is no"
  },
  {
    "startMs": 694480,
    "text": "single answer I'm coming to. Uh there is"
  },
  {
    "startMs": 696160,
    "text": "no uh no no uh no single solution. This"
  },
  {
    "startMs": 699440,
    "text": "is uh another work we did earlier. This"
  },
  {
    "startMs": 701600,
    "text": "is all before skilled learning from"
  },
  {
    "startMs": 703839,
    "text": "human videos. You watch the human do"
  },
  {
    "startMs": 705600,
    "text": "things and then robot copies this. Again"
  },
  {
    "startMs": 707680,
    "text": "high diversity. You can use videos from"
  },
  {
    "startMs": 709519,
    "text": "YouTube etc. Highly scalable but it's a"
  },
  {
    "startMs": 712320,
    "text": "very poor form of data because it's very"
  },
  {
    "startMs": 715120,
    "text": "far from robot. Okay. So what is the"
  },
  {
    "startMs": 718320,
    "text": "solution here? The answer is there is no"
  },
  {
    "startMs": 720079,
    "text": "there is no golden path. You have to"
  },
  {
    "startMs": 722240,
    "text": "think about data in the context of"
  },
  {
    "startMs": 725680,
    "text": "different features. And in my opinion"
  },
  {
    "startMs": 728160,
    "text": "there are only three features that"
  },
  {
    "startMs": 729760,
    "text": "matter in in robotics."
  },
  {
    "startMs": 732240,
    "text": "scalability, diversity and how close you"
  },
  {
    "startMs": 735360,
    "text": "are to your robot uh uh robot joint"
  },
  {
    "startMs": 737839,
    "text": "angles. Okay, scalability means can I"
  },
  {
    "startMs": 740480,
    "text": "quickly scale it across scenarios."
  },
  {
    "startMs": 742399,
    "text": "Diversity means can I get diverse data"
  },
  {
    "startMs": 745279,
    "text": "because just having 100 trillion tokens"
  },
  {
    "startMs": 747440,
    "text": "is completely useless if they're all in"
  },
  {
    "startMs": 750000,
    "text": "the same environment and same scenarios."
  },
  {
    "startMs": 751839,
    "text": "Okay. And closeness to robot mean how"
  },
  {
    "startMs": 754720,
    "text": "far are you from the ground truth of"
  },
  {
    "startMs": 756560,
    "text": "robot own joint angles. So if you look"
  },
  {
    "startMs": 758639,
    "text": "at simulation very scalable low diverse"
  },
  {
    "startMs": 762000,
    "text": "uh uh but moderately close to robot"
  },
  {
    "startMs": 764160,
    "text": "human videos are very scalable and"
  },
  {
    "startMs": 765680,
    "text": "diverse but very far from robot data you"
  },
  {
    "startMs": 767519,
    "text": "have to learn to map the human to robot"
  },
  {
    "startMs": 769839,
    "text": "this other two teleop and the and the"
  },
  {
    "startMs": 771920,
    "text": "manipulation interfaces they are kind of"
  },
  {
    "startMs": 773600,
    "text": "in between and you can see here uh right"
  },
  {
    "startMs": 776240,
    "text": "now around the world if you look at"
  },
  {
    "startMs": 777600,
    "text": "different companies they're all focusing"
  },
  {
    "startMs": 779360,
    "text": "on one of these approaches most of them"
  },
  {
    "startMs": 782480,
    "text": "are on teleop or or yumi setups very few"
  },
  {
    "startMs": 785600,
    "text": "on everything else and but there There"
  },
  {
    "startMs": 787839,
    "text": "is no golden answer like they every each"
  },
  {
    "startMs": 789839,
    "text": "one of them have a downside but the"
  },
  {
    "startMs": 791920,
    "text": "golden light here is that they are all"
  },
  {
    "startMs": 794560,
    "text": "complemented to each other like they are"
  },
  {
    "startMs": 796959,
    "text": "not their cons are not exactly matching"
  },
  {
    "startMs": 799519,
    "text": "from each other. So this is where uh the"
  },
  {
    "startMs": 801600,
    "text": "recipe that we have converged to over"
  },
  {
    "startMs": 803279,
    "text": "the years is to realize"
  },
  {
    "startMs": 806240,
    "text": "separate the training into two parts"
  },
  {
    "startMs": 807920,
    "text": "pre-training and post- training but"
  },
  {
    "startMs": 809200,
    "text": "that's not surprise right but how do we"
  },
  {
    "startMs": 810959,
    "text": "pre-train you want to pre-train using"
  },
  {
    "startMs": 812480,
    "text": "data which is highly scalable and"
  },
  {
    "startMs": 813839,
    "text": "diverse but maybe low quality so"
  },
  {
    "startMs": 815839,
    "text": "simulation human videos etc. So that's"
  },
  {
    "startMs": 818160,
    "text": "how you pre-train and then for post"
  },
  {
    "startMs": 820320,
    "text": "training you use data from telly"
  },
  {
    "startMs": 821920,
    "text": "operation. Now what is teleoperation"
  },
  {
    "startMs": 823680,
    "text": "data? It's very high quality but low in"
  },
  {
    "startMs": 825760,
    "text": "amount. Okay, high quality low in"
  },
  {
    "startMs": 828079,
    "text": "amount. It exactly reminds us of the"
  },
  {
    "startMs": 830079,
    "text": "recipe in language models. You pre-train"
  },
  {
    "startMs": 832639,
    "text": "data on the internet then you post train"
  },
  {
    "startMs": 834560,
    "text": "for coding for uh uh for your own"
  },
  {
    "startMs": 837120,
    "text": "company etc."
  },
  {
    "startMs": 838959,
    "text": "But in robotics there is one more bucket"
  },
  {
    "startMs": 840880,
    "text": "which is deployment data and deployment"
  },
  {
    "startMs": 843519,
    "text": "data is highly scalable once deployment"
  },
  {
    "startMs": 846480,
    "text": "scaled. So over time this data will take"
  },
  {
    "startMs": 849839,
    "text": "over everything else. And what we are"
  },
  {
    "startMs": 852320,
    "text": "trying to build is what we call this"
  },
  {
    "startMs": 854639,
    "text": "data flywheel which goes which takes"
  },
  {
    "startMs": 857040,
    "text": "this data from post training time and"
  },
  {
    "startMs": 858560,
    "text": "puts back in pre-training. And now you"
  },
  {
    "startMs": 860560,
    "text": "can see why this idea of omniodied brain"
  },
  {
    "startMs": 863199,
    "text": "is extremely important because this is"
  },
  {
    "startMs": 865519,
    "text": "very unlikely that you have only one"
  },
  {
    "startMs": 867680,
    "text": "robot only one version deployed forever"
  },
  {
    "startMs": 870800,
    "text": "in every task around the world. This"
  },
  {
    "startMs": 872800,
    "text": "never happens in any area of hardware"
  },
  {
    "startMs": 875519,
    "text": "except chips because they are very hard"
  },
  {
    "startMs": 877040,
    "text": "to manufacture. Uh and you can still see"
  },
  {
    "startMs": 879279,
    "text": "even in the chips uh inference chips are"
  },
  {
    "startMs": 881760,
    "text": "coming left and right for many companies"
  },
  {
    "startMs": 883199,
    "text": "these days. So in hardware this is never"
  },
  {
    "startMs": 885279,
    "text": "the case. you have only one version of"
  },
  {
    "startMs": 886720,
    "text": "the robot, one shape, which is why"
  },
  {
    "startMs": 888639,
    "text": "omniodied intelligence is the enabler of"
  },
  {
    "startMs": 892000,
    "text": "what we call a deployment data flywheel."
  },
  {
    "startMs": 894720,
    "text": "Okay, so let's now look at a few results"
  },
  {
    "startMs": 897920,
    "text": "uh very quick. Now this brain is"
  },
  {
    "startMs": 900240,
    "text": "extremely general. So we can do variety"
  },
  {
    "startMs": 902000,
    "text": "of tasks very quickly. You may have you"
  },
  {
    "startMs": 904639,
    "text": "have se you may have seen many tasks"
  },
  {
    "startMs": 905920,
    "text": "like laundry folding etc. It's very"
  },
  {
    "startMs": 907680,
    "text": "popular task in Silicon Valley for some"
  },
  {
    "startMs": 909279,
    "text": "reason. uh and uh and the and the"
  },
  {
    "startMs": 911839,
    "text": "argument here is uh when have you ever"
  },
  {
    "startMs": 914959,
    "text": "thought while folding a t-shirt that if"
  },
  {
    "startMs": 917760,
    "text": "I miss my hand by 1 cm my t-shirt fold"
  },
  {
    "startMs": 921120,
    "text": "will be a blunder you don't think like"
  },
  {
    "startMs": 923680,
    "text": "this people don't even think while"
  },
  {
    "startMs": 924959,
    "text": "folding they just hold anywhere you just"
  },
  {
    "startMs": 926560,
    "text": "do something so that the tolerance for"
  },
  {
    "startMs": 928639,
    "text": "error is extremely extremely high then"
  },
  {
    "startMs": 932000,
    "text": "why is this task hard"
  },
  {
    "startMs": 934959,
    "text": "anybody why why do people get why do"
  },
  {
    "startMs": 937440,
    "text": "people get fooled into believing this"
  },
  {
    "startMs": 938720,
    "text": "task is"
  },
  {
    "startMs": 939760,
    "text": "Let me put it that way."
  },
  {
    "startMs": 942399,
    "text": "Fabric, right? Why is fabric hard?"
  },
  {
    "startMs": 946399,
    "text": "Simulation. But nobody's using"
  },
  {
    "startMs": 947519,
    "text": "simulation anyway. This is all from"
  },
  {
    "startMs": 948399,
    "text": "telly operation. Why is this hard? You"
  },
  {
    "startMs": 951199,
    "text": "know why is it hard? Because it is hard"
  },
  {
    "startMs": 953680,
    "text": "for classical version of robotics."
  },
  {
    "startMs": 956480,
    "text": "Classically in robotics, people would"
  },
  {
    "startMs": 958000,
    "text": "model the whole physics, create models"
  },
  {
    "startMs": 960079,
    "text": "by hand and then do this. It's very hard"
  },
  {
    "startMs": 961839,
    "text": "for that. But for deep learning based"
  },
  {
    "startMs": 963680,
    "text": "robotics, this is the easiest task"
  },
  {
    "startMs": 965360,
    "text": "possible because it has high tolerance"
  },
  {
    "startMs": 967279,
    "text": "for error. So it is completely uh now so"
  },
  {
    "startMs": 970560,
    "text": "many companies focusing on this task."
  },
  {
    "startMs": 972320,
    "text": "Now what is hard is I would say"
  },
  {
    "startMs": 973680,
    "text": "something about like let's say this"
  },
  {
    "startMs": 974800,
    "text": "task."
  },
  {
    "startMs": 977040,
    "text": "If I ask you before seeing this video is"
  },
  {
    "startMs": 979360,
    "text": "this task doable without having hands"
  },
  {
    "startMs": 982240,
    "text": "very likely half the people will say no"
  },
  {
    "startMs": 984639,
    "text": "because it requires putting like how"
  },
  {
    "startMs": 986399,
    "text": "many of you have lost airports? Uh like"
  },
  {
    "startMs": 989199,
    "text": "and in our company there's a channel"
  },
  {
    "startMs": 991040,
    "text": "called right airpod because people keep"
  },
  {
    "startMs": 992720,
    "text": "losing their right. Now in this case the"
  },
  {
    "startMs": 995519,
    "text": "robot does not have hand it has gripper."
  },
  {
    "startMs": 997600,
    "text": "So here the task is very hard for the"
  },
  {
    "startMs": 1000320,
    "text": "gripper. So it requires a higher level"
  },
  {
    "startMs": 1002720,
    "text": "of intelligence because the arm has to"
  },
  {
    "startMs": 1004639,
    "text": "go and orient itself to pick up the"
  },
  {
    "startMs": 1007600,
    "text": "airpod in the right manner such that it"
  },
  {
    "startMs": 1010639,
    "text": "can be inserted because the grippers can"
  },
  {
    "startMs": 1012639,
    "text": "only close up and like like this."
  },
  {
    "startMs": 1014399,
    "text": "They're parallejo grippers. So you"
  },
  {
    "startMs": 1016079,
    "text": "cannot turn the uh the the the airpod at"
  },
  {
    "startMs": 1019279,
    "text": "the very end. Now what you are seeing"
  },
  {
    "startMs": 1020959,
    "text": "here these are not real airpods. They"
  },
  {
    "startMs": 1022560,
    "text": "are fake ones from teu uh like $5 $10"
  },
  {
    "startMs": 1025520,
    "text": "each. So they don't have a magnet"
  },
  {
    "startMs": 1027120,
    "text": "inside. So the robot has to really work"
  },
  {
    "startMs": 1029199,
    "text": "hard to put this inside properly because"
  },
  {
    "startMs": 1031439,
    "text": "there is no magnet to pull it uh easily."
  },
  {
    "startMs": 1033600,
    "text": "So these are all fake uh fake ones. This"
  },
  {
    "startMs": 1035520,
    "text": "is even harder than it appears uh in the"
  },
  {
    "startMs": 1037360,
    "text": "video. You can also uh like once you"
  },
  {
    "startMs": 1041760,
    "text": "build the general brain behind the"
  },
  {
    "startMs": 1043039,
    "text": "scene, you can also go and learn it from"
  },
  {
    "startMs": 1045039,
    "text": "variety of just human videos without"
  },
  {
    "startMs": 1046720,
    "text": "actually having any finetuning data at"
  },
  {
    "startMs": 1048240,
    "text": "tell operation time. So in this scenario"
  },
  {
    "startMs": 1050080,
    "text": "what we did, we trained on human videos"
  },
  {
    "startMs": 1052880,
    "text": "like this like egocentric videos. we are"
  },
  {
    "startMs": 1054799,
    "text": "now trying to transfer it to more third"
  },
  {
    "startMs": 1056320,
    "text": "person uh uh videos and if third person"
  },
  {
    "startMs": 1059280,
    "text": "works you can learn from YouTube any any"
  },
  {
    "startMs": 1061280,
    "text": "kind of open source data. So in this"
  },
  {
    "startMs": 1063840,
    "text": "case we see the video and then we add"
  },
  {
    "startMs": 1066080,
    "text": "less than 1 hour of robot data. So very"
  },
  {
    "startMs": 1068240,
    "text": "quick uh transformation and then the it"
  },
  {
    "startMs": 1070880,
    "text": "can transfer to humanoids. Now these"
  },
  {
    "startMs": 1073440,
    "text": "ones have hands. Hands or no hands is a"
  },
  {
    "startMs": 1076320,
    "text": "big debate. People often use hands as an"
  },
  {
    "startMs": 1078640,
    "text": "excuse as to why robotics is not here"
  },
  {
    "startMs": 1081440,
    "text": "but that's not the case. It's always the"
  },
  {
    "startMs": 1083360,
    "text": "intelligence uh behind the scene."
  },
  {
    "startMs": 1087440,
    "text": "We don't deploy hands right now because"
  },
  {
    "startMs": 1089039,
    "text": "there are none available which can be"
  },
  {
    "startMs": 1090960,
    "text": "deployed in factories. They all break"
  },
  {
    "startMs": 1092799,
    "text": "within 100 yards. Uh pick anyone."
  },
  {
    "startMs": 1097679,
    "text": "If you have better hand uh I would love"
  },
  {
    "startMs": 1099520,
    "text": "to buy 10 units ASAP to test. So it's"
  },
  {
    "startMs": 1103360,
    "text": "robust to scenarios. Now the idea is you"
  },
  {
    "startMs": 1106720,
    "text": "can do many tasks by watching humans."
  },
  {
    "startMs": 1108400,
    "text": "And the reason we can work with very"
  },
  {
    "startMs": 1110640,
    "text": "less data which is less than one hour of"
  },
  {
    "startMs": 1112320,
    "text": "robot data is because robot imagines"
  },
  {
    "startMs": 1114799,
    "text": "things in its own head and tries to"
  },
  {
    "startMs": 1116799,
    "text": "multiply the learning from for many"
  },
  {
    "startMs": 1118640,
    "text": "scenarios. So what you see here is a"
  },
  {
    "startMs": 1120320,
    "text": "completely fake video uh you can call it"
  },
  {
    "startMs": 1122880,
    "text": "robot's dream. So it's inside the"
  },
  {
    "startMs": 1124799,
    "text": "robot's own model where it can imagine"
  },
  {
    "startMs": 1126720,
    "text": "scenarios uh in so this is not real"
  },
  {
    "startMs": 1128960,
    "text": "data. This is all fake uh from own"
  },
  {
    "startMs": 1131360,
    "text": "robot's own model. You can transfer it"
  },
  {
    "startMs": 1133679,
    "text": "to you know more complex task and even"
  },
  {
    "startMs": 1136400,
    "text": "lower cost hardware. So this is a egg uh"
  },
  {
    "startMs": 1139520,
    "text": "uh egg making like omelette making task."
  },
  {
    "startMs": 1142080,
    "text": "Uh now here the these this whole setup"
  },
  {
    "startMs": 1145200,
    "text": "costs about $4,000. So extremely cheap"
  },
  {
    "startMs": 1147600,
    "text": "arms uh compared to the uh what you see"
  },
  {
    "startMs": 1150480,
    "text": "out there. And if you notice here this"
  },
  {
    "startMs": 1154080,
    "text": "setup is too cheap to even have any"
  },
  {
    "startMs": 1156320,
    "text": "sensors. So the only sensor here is just"
  },
  {
    "startMs": 1158400,
    "text": "a camera. No force nothing else. And the"
  },
  {
    "startMs": 1160880,
    "text": "robot can do task which require forces"
  },
  {
    "startMs": 1163919,
    "text": "uh from vision. Now I'm not saying"
  },
  {
    "startMs": 1165760,
    "text": "that's a future like you should not be"
  },
  {
    "startMs": 1167520,
    "text": "doing this. I'm sure the sensors will"
  },
  {
    "startMs": 1169120,
    "text": "improve but"
  },
  {
    "startMs": 1171200,
    "text": ">> from the existing sensors we are way"
  },
  {
    "startMs": 1173679,
    "text": "behind than where we can be from just"
  },
  {
    "startMs": 1176000,
    "text": "intelligence perspective."
  },
  {
    "startMs": 1177679,
    "text": ">> Okay. So this can keep on going. This is"
  },
  {
    "startMs": 1180400,
    "text": "my adviser from Berkeley. He did not"
  },
  {
    "startMs": 1182559,
    "text": "believe the robots can do it. He just"
  },
  {
    "startMs": 1184000,
    "text": "kept standing for like uh 1 hour and the"
  },
  {
    "startMs": 1186080,
    "text": "robot kept making omelette. Uh and and"
  },
  {
    "startMs": 1188480,
    "text": "the interesting part here is this is"
  },
  {
    "startMs": 1190000,
    "text": "fully end to end system. There is no"
  },
  {
    "startMs": 1191919,
    "text": "state machine nothing. So the reason"
  },
  {
    "startMs": 1194000,
    "text": "robot is looking egg because there is an"
  },
  {
    "startMs": 1195679,
    "text": "empty plate in the front. I don't know"
  },
  {
    "startMs": 1197120,
    "text": "why it's fluctuating. It's not in the"
  },
  {
    "startMs": 1198559,
    "text": "there's no cut in the video. I I assure"
  },
  {
    "startMs": 1200799,
    "text": "you of that. It's HDMI problem. So as"
  },
  {
    "startMs": 1203440,
    "text": "soon as you remove the plate uh sorry I"
  },
  {
    "startMs": 1206480,
    "text": "don't know why this is happening."
  },
  {
    "startMs": 1208880,
    "text": "Yeah. So there's no state machine when"
  },
  {
    "startMs": 1210480,
    "text": "the robot starts when the robot ends."
  },
  {
    "startMs": 1211840,
    "text": "It's all automated and it's very robust"
  },
  {
    "startMs": 1213840,
    "text": "to even disturbances. You can change"
  },
  {
    "startMs": 1215520,
    "text": "objects around. Uh you can add these are"
  },
  {
    "startMs": 1218320,
    "text": "all unseen and unseen objects. Uh only"
  },
  {
    "startMs": 1220720,
    "text": "the pan and the gas stove is same."
  },
  {
    "startMs": 1222400,
    "text": "Everything else is different and the"
  },
  {
    "startMs": 1224240,
    "text": "robot can keep on going. So you can get"
  },
  {
    "startMs": 1226000,
    "text": "basic robustness. So this was this is"
  },
  {
    "startMs": 1228240,
    "text": "all trained with less than 10 hours of"
  },
  {
    "startMs": 1230000,
    "text": "data. Uh so it's very low data to be"
  },
  {
    "startMs": 1232559,
    "text": "learning this robustness and it's coming"
  },
  {
    "startMs": 1234080,
    "text": "from the base model behind the scene."
  },
  {
    "startMs": 1237360,
    "text": "I'm skipping uh this in the interest of"
  },
  {
    "startMs": 1239200,
    "text": "time. So unlike traditional robotics"
  },
  {
    "startMs": 1242080,
    "text": "pipelines where you have planning,"
  },
  {
    "startMs": 1243520,
    "text": "mapping etc. This is an end toend brain."
  },
  {
    "startMs": 1246000,
    "text": "Now end to end is heavily used term in"
  },
  {
    "startMs": 1249039,
    "text": "in many areas of AI. But when I mean end"
  },
  {
    "startMs": 1251919,
    "text": "to end, I really mean end to end. It"
  },
  {
    "startMs": 1253919,
    "text": "reads directly from the cameras and it"
  },
  {
    "startMs": 1255919,
    "text": "applies power directly to motors. So we"
  },
  {
    "startMs": 1259360,
    "text": "use nothing uh in between except just a"
  },
  {
    "startMs": 1262240,
    "text": "P at the very end to go from 100 to 500"
  },
  {
    "startMs": 1264400,
    "text": "Hz. But P is not robotics contribution."
  },
  {
    "startMs": 1266799,
    "text": "This is before it predates to like World"
  },
  {
    "startMs": 1268400,
    "text": "War II or something. Okay. Now the the"
  },
  {
    "startMs": 1271440,
    "text": "the interesting part is vision for us is"
  },
  {
    "startMs": 1274000,
    "text": "yet another input. So nothing that crazy"
  },
  {
    "startMs": 1276720,
    "text": "about it. So if you have seen, you may"
  },
  {
    "startMs": 1279039,
    "text": "have seen a lot of lot of results of"
  },
  {
    "startMs": 1280799,
    "text": "robots dancing, doing karate, kung fu,"
  },
  {
    "startMs": 1282960,
    "text": "backflip."
  },
  {
    "startMs": 1284960,
    "text": "Go back and think how many results have"
  },
  {
    "startMs": 1287440,
    "text": "you seen of robots going up and down"
  },
  {
    "startMs": 1289039,
    "text": "stairs. Very few. Like even the top"
  },
  {
    "startMs": 1292000,
    "text": "companies out there in humanoid, they"
  },
  {
    "startMs": 1294240,
    "text": "have not shown beyond one sample stair"
  },
  {
    "startMs": 1297120,
    "text": "just to check the tick tick box. You do."
  },
  {
    "startMs": 1299840,
    "text": "And people talk about humanoids are"
  },
  {
    "startMs": 1301600,
    "text": "coming, China versus US, all of this is"
  },
  {
    "startMs": 1304240,
    "text": "sort of BS. If humanoids cannot climb"
  },
  {
    "startMs": 1306559,
    "text": "stairs, what's the point of having legs?"
  },
  {
    "startMs": 1309760,
    "text": "All right, that's the only reason why"
  },
  {
    "startMs": 1310960,
    "text": "you have legs. Now, this is actually a"
  },
  {
    "startMs": 1312799,
    "text": "Moex paradox add action. Left one"
  },
  {
    "startMs": 1316480,
    "text": "actually is much easier for a robot. A"
  },
  {
    "startMs": 1318320,
    "text": "backflip, dancing, etc. is much easier"
  },
  {
    "startMs": 1320400,
    "text": "for the robot. And you may think, okay,"
  },
  {
    "startMs": 1322320,
    "text": "why is this paradox exist? Think one"
  },
  {
    "startMs": 1324720,
    "text": "level deeper. When a robot is doing a"
  },
  {
    "startMs": 1327600,
    "text": "back flip, it has to only know about its"
  },
  {
    "startMs": 1330880,
    "text": "own body, nothing else. Right? So it's"
  },
  {
    "startMs": 1334559,
    "text": "and when everything is known or fully"
  },
  {
    "startMs": 1336960,
    "text": "observed that's what computers are good"
  },
  {
    "startMs": 1339360,
    "text": "at because they can do they can simulate"
  },
  {
    "startMs": 1342640,
    "text": "every possible uh uh setup but the right"
  },
  {
    "startMs": 1345679,
    "text": "one even simply climbing on stairs"
  },
  {
    "startMs": 1347919,
    "text": "requires seeing the stair at the first"
  },
  {
    "startMs": 1349760,
    "text": "at the first place or what's the height"
  },
  {
    "startMs": 1351600,
    "text": "what's the width you don't measure"
  },
  {
    "startMs": 1353280,
    "text": "height and width exactly but you adjust"
  },
  {
    "startMs": 1354880,
    "text": "to all the disturbances and that"
  },
  {
    "startMs": 1356080,
    "text": "requires vision so right one requires"
  },
  {
    "startMs": 1358000,
    "text": "understanding and that's why it is hard"
  },
  {
    "startMs": 1359760,
    "text": "and you don't see any of this but for us"
  },
  {
    "startMs": 1361200,
    "text": "it's just get another result it doesn't"
  },
  {
    "startMs": 1362799,
    "text": "really matter so we And uh we put this"
  },
  {
    "startMs": 1364880,
    "text": "result out one and a half year ago. We"
  },
  {
    "startMs": 1366640,
    "text": "shot this like two and a half years ago."
  },
  {
    "startMs": 1368480,
    "text": "Uh but uh like this robot can go any"
  },
  {
    "startMs": 1371600,
    "text": "scenario. It is not stumbling on things."
  },
  {
    "startMs": 1373919,
    "text": "It intentionally steps over things and"
  },
  {
    "startMs": 1375520,
    "text": "there is no mapping or planning. It"
  },
  {
    "startMs": 1377200,
    "text": "doesn't make any 3D map of the"
  },
  {
    "startMs": 1378720,
    "text": "surroundings. It's all operating from"
  },
  {
    "startMs": 1380880,
    "text": "camera on the torso. And you can put"
  },
  {
    "startMs": 1382720,
    "text": "this in any kind of setup, any kind of"
  },
  {
    "startMs": 1384559,
    "text": "stairs. Uh I'm going faster here. These"
  },
  {
    "startMs": 1387679,
    "text": "are, you know, uh fire escape stairs."
  },
  {
    "startMs": 1391120,
    "text": "It's a bit weird. Fire escape stairs are"
  },
  {
    "startMs": 1393440,
    "text": "to be used when there is fire in urgency"
  },
  {
    "startMs": 1395280,
    "text": "and they are the hardest in every"
  },
  {
    "startMs": 1396880,
    "text": "building. Uh so we are testing in all in"
  },
  {
    "startMs": 1399760,
    "text": "all those setups but you can put this in"
  },
  {
    "startMs": 1401360,
    "text": "anywhere you can disturb it and stay"
  },
  {
    "startMs": 1402799,
    "text": "doesn't really matter. This is"
  },
  {
    "startMs": 1403840,
    "text": "superhuman capability because robot"
  },
  {
    "startMs": 1405760,
    "text": "cannot see it's being pulled. So it's a"
  },
  {
    "startMs": 1408240,
    "text": "surprise factor for the robot that"
  },
  {
    "startMs": 1409600,
    "text": "you're pulling the violet while"
  },
  {
    "startMs": 1410799,
    "text": "climbing. And again it's the same model"
  },
  {
    "startMs": 1412559,
    "text": "everywhere in all these scenarios. So"
  },
  {
    "startMs": 1415360,
    "text": "very easy although easy parkour does"
  },
  {
    "startMs": 1417840,
    "text": "look fun uh uh does look fun and people"
  },
  {
    "startMs": 1420880,
    "text": "often find like uh this is uh you know"
  },
  {
    "startMs": 1425520,
    "text": "the robots can do all this but this is"
  },
  {
    "startMs": 1428320,
    "text": "so much easier than what I showed"
  },
  {
    "startMs": 1430799,
    "text": "earlier and even if you see these videos"
  },
  {
    "startMs": 1433280,
    "text": "right now many of you will find this"
  },
  {
    "startMs": 1435120,
    "text": "more impressive even though I'm telling"
  },
  {
    "startMs": 1436480,
    "text": "you this is easy it takes half an hour"
  },
  {
    "startMs": 1438480,
    "text": "to train this while the previous one it"
  },
  {
    "startMs": 1440000,
    "text": "takes much longer uh and much more"
  },
  {
    "startMs": 1441440,
    "text": "difficult to train"
  },
  {
    "startMs": 1443440,
    "text": "so you can take this Based model and you"
  },
  {
    "startMs": 1445760,
    "text": "can transfer to variety of tasks very"
  },
  {
    "startMs": 1447520,
    "text": "quickly. So here this is a very old"
  },
  {
    "startMs": 1449600,
    "text": "result from one and a half year ago uh"
  },
  {
    "startMs": 1451520,
    "text": "for inserting LAN cables etc. We have"
  },
  {
    "startMs": 1453440,
    "text": "very advanced systems now but we are"
  },
  {
    "startMs": 1455200,
    "text": "deploying these models already across"
  },
  {
    "startMs": 1457360,
    "text": "variety of application. So for instance"
  },
  {
    "startMs": 1459520,
    "text": "because to to build the data flywhe the"
  },
  {
    "startMs": 1461679,
    "text": "hardest part is the deployment itself."
  },
  {
    "startMs": 1464080,
    "text": "There are so many other issues beside"
  },
  {
    "startMs": 1465919,
    "text": "intelligence and hardware that come up"
  },
  {
    "startMs": 1467760,
    "text": "when you deploy robotics. It is not the"
  },
  {
    "startMs": 1469840,
    "text": "same. This is not same same as deploying"
  },
  {
    "startMs": 1472640,
    "text": "chat GP over an app because you have to"
  },
  {
    "startMs": 1475200,
    "text": "work around many of the and I think"
  },
  {
    "startMs": 1476880,
    "text": "Skyio gave a talk before this and they"
  },
  {
    "startMs": 1479120,
    "text": "can talk to you about all day about the"
  },
  {
    "startMs": 1480720,
    "text": "hardness of deployment. So we have to"
  },
  {
    "startMs": 1482880,
    "text": "start deployment now so that we can have"
  },
  {
    "startMs": 1485200,
    "text": "this data flywheel in a foreseeable"
  },
  {
    "startMs": 1486720,
    "text": "future and I'll give you a few examples"
  },
  {
    "startMs": 1488559,
    "text": "of deployment we are doing. So one of"
  },
  {
    "startMs": 1490000,
    "text": "the examples with Nvidia like Nvidia is"
  },
  {
    "startMs": 1492159,
    "text": "opening their first factory in Houston."
  },
  {
    "startMs": 1494320,
    "text": "uh and their GPUs what you use right now"
  },
  {
    "startMs": 1496640,
    "text": "they're all built outside US in in"
  },
  {
    "startMs": 1498480,
    "text": "Taiwan mainly and it's all done manually"
  },
  {
    "startMs": 1500799,
    "text": "over there and uh humos are really"
  },
  {
    "startMs": 1502799,
    "text": "efficient and really uh really good at"
  },
  {
    "startMs": 1504640,
    "text": "making these things but sustaining cost"
  },
  {
    "startMs": 1507039,
    "text": "and scalability and throughput here in"
  },
  {
    "startMs": 1508880,
    "text": "US it's very hard to maintain without"
  },
  {
    "startMs": 1510480,
    "text": "this labor force so here we are"
  },
  {
    "startMs": 1512880,
    "text": "automating uh this GPU assembly for them"
  },
  {
    "startMs": 1515919,
    "text": "this was we did a live demo in Nvidia"
  },
  {
    "startMs": 1518159,
    "text": "GTC this is deployed in factory already"
  },
  {
    "startMs": 1520799,
    "text": "uh last week so it's already live but"
  },
  {
    "startMs": 1523039,
    "text": "the interesting part I want to highlight"
  },
  {
    "startMs": 1525919,
    "text": "this is a very traditional factory task"
  },
  {
    "startMs": 1528960,
    "text": "but if you look at the the whole setup"
  },
  {
    "startMs": 1532960,
    "text": "this is extremely randomized extremely"
  },
  {
    "startMs": 1534960,
    "text": "noisy and you go to any factory"
  },
  {
    "startMs": 1537679,
    "text": "traditional setup they're extremely"
  },
  {
    "startMs": 1539039,
    "text": "clean so here the robot the same brain"
  },
  {
    "startMs": 1542480,
    "text": "can not only do very precise task it can"
  },
  {
    "startMs": 1544720,
    "text": "be very robust to any kind of"
  },
  {
    "startMs": 1546000,
    "text": "disturbance which means now these robots"
  },
  {
    "startMs": 1548559,
    "text": "do not need to be in a cage uh and they"
  },
  {
    "startMs": 1551120,
    "text": "can be uh just deployed as is on on"
  },
  {
    "startMs": 1553840,
    "text": "these factory lines without any change"
  },
  {
    "startMs": 1555679,
    "text": "to any factory line alongside humans and"
  },
  {
    "startMs": 1558159,
    "text": "wherever there is humans there is mess."
  },
  {
    "startMs": 1560000,
    "text": "Okay, you can put the same brain to"
  },
  {
    "startMs": 1561679,
    "text": "delivery applications. So here the robot"
  },
  {
    "startMs": 1563520,
    "text": "delivers from the from the truck to the"
  },
  {
    "startMs": 1565520,
    "text": "to the door. It has to find where the"
  },
  {
    "startMs": 1567760,
    "text": "front door is. So it's a common sense"
  },
  {
    "startMs": 1569039,
    "text": "problem. It's not mobility problem. And"
  },
  {
    "startMs": 1570640,
    "text": "we already have been delivering uh"
  },
  {
    "startMs": 1572240,
    "text": "packages to with working behind the"
  },
  {
    "startMs": 1574400,
    "text": "scene with many partners to delivering"
  },
  {
    "startMs": 1576000,
    "text": "packages to front door of the houses uh"
  },
  {
    "startMs": 1577919,
    "text": "in these areas. same setup works across"
  },
  {
    "startMs": 1580640,
    "text": "warehouses and all this. Um now just to"
  },
  {
    "startMs": 1584960,
    "text": "close the talk uh close the topic I"
  },
  {
    "startMs": 1587200,
    "text": "mentioned this is an omnibody brain and"
  },
  {
    "startMs": 1589120,
    "text": "I hope it's very clear why it's"
  },
  {
    "startMs": 1590640,
    "text": "extremely essential to have an omnibody"
  },
  {
    "startMs": 1592320,
    "text": "brain to build data flywheel but there's"
  },
  {
    "startMs": 1594799,
    "text": "one extra benefit and the benefit is"
  },
  {
    "startMs": 1596880,
    "text": "your robots may break over time and uh"
  },
  {
    "startMs": 1599440,
    "text": "or the safety uh scenarios and safety is"
  },
  {
    "startMs": 1602159,
    "text": "a byproduct of this omnibody brain"
  },
  {
    "startMs": 1604320,
    "text": "because if you have a we can put the I'm"
  },
  {
    "startMs": 1606400,
    "text": "just skipping very fast here so you can"
  },
  {
    "startMs": 1607600,
    "text": "see the online this is all online uh but"
  },
  {
    "startMs": 1609919,
    "text": "we can put the same brain across each of"
  },
  {
    "startMs": 1611760,
    "text": "these systems each of these robots"
  },
  {
    "startMs": 1615919,
    "text": "and in this case we did not even train"
  },
  {
    "startMs": 1617840,
    "text": "on these robots. This is completely zero"
  },
  {
    "startMs": 1619440,
    "text": "shot transfer to all these humanoids and"
  },
  {
    "startMs": 1621279,
    "text": "all these things and even if the robot"
  },
  {
    "startMs": 1623440,
    "text": "breaks like here the leg breaks it can"
  },
  {
    "startMs": 1626159,
    "text": "recover within milliseconds because when"
  },
  {
    "startMs": 1628000,
    "text": "your leg breaks your three-legged robot"
  },
  {
    "startMs": 1630640,
    "text": "is a new robot. So it doesn't really"
  },
  {
    "startMs": 1632880,
    "text": "matter what's the shape of the robot is"
  },
  {
    "startMs": 1634960,
    "text": "anything changing here we disabled the"
  },
  {
    "startMs": 1636400,
    "text": "legs of the robot it learns to learn it"
  },
  {
    "startMs": 1638480,
    "text": "learns to walk on two legs in just three"
  },
  {
    "startMs": 1640080,
    "text": "trials. So what you are seeing on screen"
  },
  {
    "startMs": 1642559,
    "text": "is all the training for the robot that"
  },
  {
    "startMs": 1644799,
    "text": "they are that that is happening in these"
  },
  {
    "startMs": 1646799,
    "text": "systems. So it's it adapts in few"
  },
  {
    "startMs": 1648720,
    "text": "milliseconds to uh uh 30 seconds or so."
  },
  {
    "startMs": 1652080,
    "text": "And like one example here it's it's"
  },
  {
    "startMs": 1653600,
    "text": "going on wheels. You jam the wheel it"
  },
  {
    "startMs": 1656640,
    "text": "starts walking. So this is a another"
  },
  {
    "startMs": 1658960,
    "text": "byproduct of omnibody brain. Safety"
  },
  {
    "startMs": 1662320,
    "text": "takes a very different meaning uh with"
  },
  {
    "startMs": 1664480,
    "text": "these kind of models. when the robot can"
  },
  {
    "startMs": 1666400,
    "text": "fly like or sorry when the robots can"
  },
  {
    "startMs": 1668320,
    "text": "walk or or operate with only half the"
  },
  {
    "startMs": 1670240,
    "text": "body. I removed some of the gory scenes"
  },
  {
    "startMs": 1672320,
    "text": "from this but we also cut the robot in"
  },
  {
    "startMs": 1673679,
    "text": "half and it can still work uh with both"
  },
  {
    "startMs": 1676080,
    "text": "the two halves separately but for that"
  },
  {
    "startMs": 1677919,
    "text": "you can go to YouTube uh and that's all"
  },
  {
    "startMs": 1679679,
    "text": "I have. Thank you."
  }
]