[
  {
    "startMs": 1309,
    "text": "[music]"
  },
  {
    "startMs": 12360,
    "text": ">> Thanks for the kind introduction. So,"
  },
  {
    "startMs": 13960,
    "text": "I'm Jason, co-founder of Diana."
  },
  {
    "startMs": 16360,
    "text": "Today, I'll talk about how we're"
  },
  {
    "startMs": 18200,
    "text": "developing high-performance and very"
  },
  {
    "startMs": 20560,
    "text": "robust generalist robotics policies."
  },
  {
    "startMs": 23720,
    "text": "Yeah."
  },
  {
    "startMs": 24600,
    "text": "So, let's jump into it. So, today's talk"
  },
  {
    "startMs": 27120,
    "text": "will focus on, you know, how we're"
  },
  {
    "startMs": 28400,
    "text": "bringing robots into commercial grade"
  },
  {
    "startMs": 30400,
    "text": "and what we are doing to make these"
  },
  {
    "startMs": 32119,
    "text": "models, like I said, very"
  },
  {
    "startMs": 33680,
    "text": "high-performance and robust. So, just a"
  },
  {
    "startMs": 35840,
    "text": "brief introduction to our company. So,"
  },
  {
    "startMs": 37720,
    "text": "our mission is to build very robust"
  },
  {
    "startMs": 40120,
    "text": "foundation model autonomy in the real"
  },
  {
    "startMs": 41760,
    "text": "world. So, we want to train models and"
  },
  {
    "startMs": 43880,
    "text": "these days also hardware to have a"
  },
  {
    "startMs": 46320,
    "text": "single platform that can do many"
  },
  {
    "startMs": 48680,
    "text": "economically useful physical tasks in"
  },
  {
    "startMs": 50800,
    "text": "the real world, like the ones we're"
  },
  {
    "startMs": 52120,
    "text": "showing here. So, the company the"
  },
  {
    "startMs": 54000,
    "text": "company was founded in September 2024."
  },
  {
    "startMs": 56520,
    "text": "We are a series A company, have raised"
  },
  {
    "startMs": 58680,
    "text": "about $120 million, and we"
  },
  {
    "startMs": 63920,
    "text": "And our thesis is that to actually bring"
  },
  {
    "startMs": 65960,
    "text": "robots into the real world, being at"
  },
  {
    "startMs": 68120,
    "text": "commercial grade doing useful tasks, the"
  },
  {
    "startMs": 70480,
    "text": "company needs to combine doing frontier"
  },
  {
    "startMs": 72560,
    "text": "research with a lot of commercial"
  },
  {
    "startMs": 74600,
    "text": "deployments, so we can build what we"
  },
  {
    "startMs": 76800,
    "text": "consider a research and deployment"
  },
  {
    "startMs": 78600,
    "text": "flywheel."
  },
  {
    "startMs": 79800,
    "text": "Uh which is that the research and"
  },
  {
    "startMs": 81520,
    "text": "hardware we do in-house, the R&D informs"
  },
  {
    "startMs": 84240,
    "text": "the kind of tasks, the kind of workflows"
  },
  {
    "startMs": 86400,
    "text": "in the real world that we can"
  },
  {
    "startMs": 87360,
    "text": "commercialize. So, we actively try to"
  },
  {
    "startMs": 89600,
    "text": "deploy. So, these days we have more than"
  },
  {
    "startMs": 91800,
    "text": "five deployment sites doing a bunch of"
  },
  {
    "startMs": 93520,
    "text": "different tasks, which I will talk about"
  },
  {
    "startMs": 95600,
    "text": "in a bit. And then by building product"
  },
  {
    "startMs": 98000,
    "text": "and by deploying robots, it does several"
  },
  {
    "startMs": 100640,
    "text": "things. One is that it can help us"
  },
  {
    "startMs": 102280,
    "text": "gather high-quality deployment data, and"
  },
  {
    "startMs": 105120,
    "text": "it can also tell us what our models and"
  },
  {
    "startMs": 107200,
    "text": "what our hardware is not good at yet."
  },
  {
    "startMs": 108880,
    "text": "So, it helps us sharpen our research"
  },
  {
    "startMs": 110480,
    "text": "focus to figure out what is the right"
  },
  {
    "startMs": 112600,
    "text": "problem to work on in robotics. Because"
  },
  {
    "startMs": 115600,
    "text": "if you're familiar with the robotics"
  },
  {
    "startMs": 117040,
    "text": "field, there are too many problems."
  },
  {
    "startMs": 118600,
    "text": "There are too many different fields you"
  },
  {
    "startMs": 120400,
    "text": "can spend your effort on. And by"
  },
  {
    "startMs": 122240,
    "text": "deploying and by building a product, we"
  },
  {
    "startMs": 124160,
    "text": "know exactly the right kind of problems"
  },
  {
    "startMs": 125960,
    "text": "that we need to focus to actually make a"
  },
  {
    "startMs": 128000,
    "text": "robotics not just a demo or videos you"
  },
  {
    "startMs": 130479,
    "text": "see on YouTube, but rather in the real"
  },
  {
    "startMs": 132440,
    "text": "world impacting millions of people's"
  },
  {
    "startMs": 134160,
    "text": "life."
  },
  {
    "startMs": 136040,
    "text": "So, before I jump into what we"
  },
  {
    "startMs": 138080,
    "text": "specifically work on at covariant, just"
  },
  {
    "startMs": 139760,
    "text": "a very quick introduction on the kind of"
  },
  {
    "startMs": 142280,
    "text": "models we're training for robotic"
  },
  {
    "startMs": 144320,
    "text": "manipulation. So, at a high level, uh we"
  },
  {
    "startMs": 147160,
    "text": "do data collection where you know,"
  },
  {
    "startMs": 148720,
    "text": "there's a lot of data being collected."
  },
  {
    "startMs": 150400,
    "text": "So, here's a video of me uh manipulating"
  },
  {
    "startMs": 153360,
    "text": "a robotic hardware to, you know, fold a"
  },
  {
    "startMs": 155280,
    "text": "t-shirt. And once you collect enough of"
  },
  {
    "startMs": 157880,
    "text": "this kind of data, you can put all of"
  },
  {
    "startMs": 160120,
    "text": "them in a large neural network. And a"
  },
  {
    "startMs": 162720,
    "text": "neural network essentially takes in a,"
  },
  {
    "startMs": 165600,
    "text": "you know, representation of the world,"
  },
  {
    "startMs": 167480,
    "text": "in this case just camera feed of what"
  },
  {
    "startMs": 169680,
    "text": "the, you know, a table looks like, and"
  },
  {
    "startMs": 171760,
    "text": "the outputs robot's, you know, joint"
  },
  {
    "startMs": 173440,
    "text": "positions or torque to actually control"
  },
  {
    "startMs": 175520,
    "text": "the robot to do the task that you have"
  },
  {
    "startMs": 177959,
    "text": "collected data on."
  },
  {
    "startMs": 179840,
    "text": "Right?"
  },
  {
    "startMs": 180959,
    "text": "So, you know, this is the high level of"
  },
  {
    "startMs": 183920,
    "text": "how you robot models in the real world"
  },
  {
    "startMs": 186320,
    "text": "function. So, how are we actually"
  },
  {
    "startMs": 188720,
    "text": "scaling this up to actually create"
  },
  {
    "startMs": 190600,
    "text": "models that's generalizable and it can"
  },
  {
    "startMs": 192320,
    "text": "do a lot of different tasks. So, at"
  },
  {
    "startMs": 194280,
    "text": "covariant we focus on what we call the"
  },
  {
    "startMs": 196080,
    "text": "pre-training data pyramid for the real"
  },
  {
    "startMs": 197880,
    "text": "world, where we gather a lot of diverse"
  },
  {
    "startMs": 200239,
    "text": "off-robot data, just because robotics"
  },
  {
    "startMs": 202320,
    "text": "data is very scarce. So, we have data"
  },
  {
    "startMs": 204600,
    "text": "captured from humans wearing cameras, uh"
  },
  {
    "startMs": 207560,
    "text": "from public data sets, and these days"
  },
  {
    "startMs": 209519,
    "text": "we're also considering some simulation"
  },
  {
    "startMs": 211239,
    "text": "data. But if you only have off-robot"
  },
  {
    "startMs": 213440,
    "text": "data, it's not actually enough to get"
  },
  {
    "startMs": 215320,
    "text": "robots to do very precise actions. So,"
  },
  {
    "startMs": 218120,
    "text": "on the robot themselves, we also collect"
  },
  {
    "startMs": 220160,
    "text": "diverse tasks collected in many"
  },
  {
    "startMs": 221959,
    "text": "different scenarios, on many different"
  },
  {
    "startMs": 224120,
    "text": "types of tasks, like industrial tasks,"
  },
  {
    "startMs": 226080,
    "text": "household, uh in the laundromat, and uh"
  },
  {
    "startMs": 228840,
    "text": "you know, hotels. And then finally,"
  },
  {
    "startMs": 230800,
    "text": "because we're also deploying robots, we"
  },
  {
    "startMs": 232640,
    "text": "can collect very high-quality deployment"
  },
  {
    "startMs": 235160,
    "text": "data, which helps the model to close"
  },
  {
    "startMs": 237640,
    "text": "train and test distribution gap. Because"
  },
  {
    "startMs": 240440,
    "text": "when you're developing robots, you know,"
  },
  {
    "startMs": 242200,
    "text": "most of the time your robots are in your"
  },
  {
    "startMs": 243960,
    "text": "facility, in your laboratories. But if"
  },
  {
    "startMs": 245920,
    "text": "you're deploying robots, it's in a very"
  },
  {
    "startMs": 247320,
    "text": "different environment. So, we found that"
  },
  {
    "startMs": 249240,
    "text": "by combining these three data sources,"
  },
  {
    "startMs": 251360,
    "text": "we can actually train large-scale"
  },
  {
    "startMs": 252800,
    "text": "foundation models that work very well in"
  },
  {
    "startMs": 254720,
    "text": "the real world. And so far, we have more"
  },
  {
    "startMs": 256920,
    "text": "than 200,000 hours of data in our"
  },
  {
    "startMs": 258920,
    "text": "training pipeline."
  },
  {
    "startMs": 260640,
    "text": "And the model architecture roughly"
  },
  {
    "startMs": 262520,
    "text": "follows a high-level reasoning model"
  },
  {
    "startMs": 265200,
    "text": "with a low-level world action model that"
  },
  {
    "startMs": 267440,
    "text": "can actually output dexterous actions at"
  },
  {
    "startMs": 270560,
    "text": "fine-grained high frequency, right? So,"
  },
  {
    "startMs": 273240,
    "text": "this is the kind of architecture we have"
  },
  {
    "startMs": 275080,
    "text": "because in the real world, if you think"
  },
  {
    "startMs": 276840,
    "text": "about a robot doing physical tasks, it"
  },
  {
    "startMs": 279280,
    "text": "needs to have a semantic understanding"
  },
  {
    "startMs": 281640,
    "text": "of the world, but also needs to"
  },
  {
    "startMs": 283200,
    "text": "understand physical interaction at a"
  },
  {
    "startMs": 285520,
    "text": "fine-grained level to be actually able"
  },
  {
    "startMs": 287360,
    "text": "to, you know, recover from mistakes and"
  },
  {
    "startMs": 289400,
    "text": "do very precise actions to complete the"
  },
  {
    "startMs": 291320,
    "text": "task."
  },
  {
    "startMs": 292480,
    "text": "So, once you combine,"
  },
  {
    "startMs": 294400,
    "text": "you know, this architecture with a lot"
  },
  {
    "startMs": 296720,
    "text": "of data, what happens is that the model"
  },
  {
    "startMs": 298840,
    "text": "can be rapidly fine-tuned to do a bunch"
  },
  {
    "startMs": 300880,
    "text": "of different tasks in the real world."
  },
  {
    "startMs": 303080,
    "text": "Yeah, so here are just a gallery of the"
  },
  {
    "startMs": 305480,
    "text": "kind of benchmark tasks we're doing in"
  },
  {
    "startMs": 307680,
    "text": "the office. So, it spans from, you know,"
  },
  {
    "startMs": 309640,
    "text": "like cleaning trash, opening a box, to,"
  },
  {
    "startMs": 312240,
    "text": "you know, things like folding towels,"
  },
  {
    "startMs": 314480,
    "text": "folding t-shirts, and also just a bunch"
  },
  {
    "startMs": 316160,
    "text": "of other tasks our researchers have"
  },
  {
    "startMs": 317760,
    "text": "thought about."
  },
  {
    "startMs": 319040,
    "text": "And what's really interesting is that"
  },
  {
    "startMs": 320640,
    "text": "with a good pre-trained model, even"
  },
  {
    "startMs": 323120,
    "text": "without any,"
  },
  {
    "startMs": 324560,
    "text": "you know, for some of these tasks I'm"
  },
  {
    "startMs": 326400,
    "text": "showing you here, they're not in the"
  },
  {
    "startMs": 328080,
    "text": "pre-training data at all. But with a"
  },
  {
    "startMs": 330080,
    "text": "good pre-training, you can actually"
  },
  {
    "startMs": 331480,
    "text": "post-train the model to do these kind of"
  },
  {
    "startMs": 333720,
    "text": "highly dexterous tool use tasks with"
  },
  {
    "startMs": 336000,
    "text": "very little amount of data. So, both of"
  },
  {
    "startMs": 338360,
    "text": "the videos you're seeing here have only"
  },
  {
    "startMs": 340400,
    "text": "been trained on with less than 1 hour of"
  },
  {
    "startMs": 342919,
    "text": "task-specific data. But you can see that"
  },
  {
    "startMs": 345360,
    "text": "the robot can repetitively"
  },
  {
    "startMs": 347440,
    "text": "do these tasks over and over without"
  },
  {
    "startMs": 349360,
    "text": "failure. And I think this is a stepping"
  },
  {
    "startMs": 351920,
    "text": "stone towards actual commercial grade"
  },
  {
    "startMs": 354280,
    "text": "real world deployment because in the"
  },
  {
    "startMs": 356120,
    "text": "real world physical tasks need to be"
  },
  {
    "startMs": 358000,
    "text": "done by humans and by robots over and"
  },
  {
    "startMs": 360440,
    "text": "over. And if your models are not"
  },
  {
    "startMs": 362240,
    "text": "reliable enough, then yes, you can shoot"
  },
  {
    "startMs": 364760,
    "text": "these kind of pretty demos, but it's"
  },
  {
    "startMs": 366840,
    "text": "still very far away from actual"
  },
  {
    "startMs": 368360,
    "text": "deployment."
  },
  {
    "startMs": 369800,
    "text": "So that actually brings us to what I"
  },
  {
    "startMs": 372000,
    "text": "consider the current status quo for"
  },
  {
    "startMs": 374800,
    "text": "training large-scale foundation models"
  },
  {
    "startMs": 376520,
    "text": "for manipulation, which is that the"
  },
  {
    "startMs": 378560,
    "text": "generous models that can do many many"
  },
  {
    "startMs": 380400,
    "text": "tasks like the ones I've been showing"
  },
  {
    "startMs": 382160,
    "text": "you. Even though they make pretty"
  },
  {
    "startMs": 383680,
    "text": "videos, but what I would tell you is"
  },
  {
    "startMs": 385520,
    "text": "that the success rate is actually"
  },
  {
    "startMs": 387640,
    "text": "not super high."
  },
  {
    "startMs": 389640,
    "text": "You know, when these models are actually"
  },
  {
    "startMs": 391120,
    "text": "deploying, they're about 80 to 90%"
  },
  {
    "startMs": 393760,
    "text": "success rate. But if you're stuck at 80"
  },
  {
    "startMs": 396000,
    "text": "to 90% success rate, then you know, the"
  },
  {
    "startMs": 398520,
    "text": "chance that you can do the same task 10"
  },
  {
    "startMs": 400919,
    "text": "times in a row is actually less than"
  },
  {
    "startMs": 402560,
    "text": "0.1%."
  },
  {
    "startMs": 404320,
    "text": "And on on the flip side, in the history"
  },
  {
    "startMs": 406600,
    "text": "of robotics, we have had very"
  },
  {
    "startMs": 408200,
    "text": "specialized robots and a very"
  },
  {
    "startMs": 409720,
    "text": "specialized machine learning pipelines"
  },
  {
    "startMs": 412080,
    "text": "to do singular tasks like pick and"
  },
  {
    "startMs": 414120,
    "text": "place. But this kind of pipeline, which"
  },
  {
    "startMs": 416520,
    "text": "is what I consider specialist models,"
  },
  {
    "startMs": 418320,
    "text": "aren't aren't very scalable, meaning"
  },
  {
    "startMs": 420120,
    "text": "that you can't take the same pipeline to"
  },
  {
    "startMs": 422080,
    "text": "just do a new task and do an arbitrary"
  },
  {
    "startMs": 424240,
    "text": "task very fast. So our mission is to"
  },
  {
    "startMs": 426680,
    "text": "resolve the status quo and you know,"
  },
  {
    "startMs": 428400,
    "text": "this kind of like dichotomy by training"
  },
  {
    "startMs": 431160,
    "text": "general purpose models that can both do"
  },
  {
    "startMs": 433000,
    "text": "many many tasks and also be reliable"
  },
  {
    "startMs": 435120,
    "text": "enough for commercial deployments."
  },
  {
    "startMs": 438120,
    "text": "So how do we actually do that? In"
  },
  {
    "startMs": 439800,
    "text": "today's talk, I'll briefly talk about"
  },
  {
    "startMs": 441480,
    "text": "some of our progress on mastering very"
  },
  {
    "startMs": 443800,
    "text": "complex tasks very reliably and also"
  },
  {
    "startMs": 446480,
    "text": "taking the same skills to be performant"
  },
  {
    "startMs": 449520,
    "text": "not only in the environments in the"
  },
  {
    "startMs": 451440,
    "text": "laboratory environments that we're"
  },
  {
    "startMs": 452520,
    "text": "training, but also being able to deploy"
  },
  {
    "startMs": 454760,
    "text": "to arbitrary customer sites without"
  },
  {
    "startMs": 456919,
    "text": "fine-tuning or without fine-tuning"
  },
  {
    "startMs": 458880,
    "text": "adaptation."
  },
  {
    "startMs": 460600,
    "text": "Okay, so let me get to this. Right, So,"
  },
  {
    "startMs": 463480,
    "text": "the first question we want to answer is"
  },
  {
    "startMs": 465280,
    "text": "can we take, you know, the general"
  },
  {
    "startMs": 467040,
    "text": "pre-training and post-training recipe"
  },
  {
    "startMs": 468560,
    "text": "that we had, but then turn these models"
  },
  {
    "startMs": 470720,
    "text": "into models that can be almost close to"
  },
  {
    "startMs": 473320,
    "text": "100% robust on uh any task. So, this is"
  },
  {
    "startMs": 477480,
    "text": "the first research result we published"
  },
  {
    "startMs": 479440,
    "text": "last year. Uh for detail, you can check"
  },
  {
    "startMs": 481520,
    "text": "out our blog post. But, at a high level,"
  },
  {
    "startMs": 484040,
    "text": "you know, if we want to make a model"
  },
  {
    "startMs": 485560,
    "text": "100% robust on many, many tasks, I think"
  },
  {
    "startMs": 488280,
    "text": "it's very good to first narrow down on a"
  },
  {
    "startMs": 490600,
    "text": "commercial use case that's reasonably"
  },
  {
    "startMs": 492600,
    "text": "hard, which allows you to make research"
  },
  {
    "startMs": 494800,
    "text": "progress, but also has commercial value."
  },
  {
    "startMs": 497360,
    "text": "So, when we first started the company uh"
  },
  {
    "startMs": 499360,
    "text": "last year, what we discovered is that if"
  },
  {
    "startMs": 501680,
    "text": "you go to any, you know, uh restaurant,"
  },
  {
    "startMs": 503720,
    "text": "you know, uh fancy restaurant or dim sum"
  },
  {
    "startMs": 505800,
    "text": "places in the US, you'll see nicely"
  },
  {
    "startMs": 507840,
    "text": "folded napkin on the table, right, for"
  },
  {
    "startMs": 509520,
    "text": "you, right? If you go to a Cheesecake"
  },
  {
    "startMs": 511000,
    "text": "Factory, you'll see those napkins. And"
  },
  {
    "startMs": 513120,
    "text": "what happens is that in the back office"
  },
  {
    "startMs": 515000,
    "text": "of the restaurant, there's usually uh"
  },
  {
    "startMs": 516760,
    "text": "workers or, you know, restaurant"
  },
  {
    "startMs": 518640,
    "text": "employees that's folding these napkins"
  },
  {
    "startMs": 520880,
    "text": "one by one by hand. And that's a very"
  },
  {
    "startMs": 523280,
    "text": "mundane process, and a lot of the"
  },
  {
    "startMs": 525480,
    "text": "customers we have talked to are looking"
  },
  {
    "startMs": 527160,
    "text": "into robots that can actually do the"
  },
  {
    "startMs": 529120,
    "text": "same task. So, this is one of the"
  },
  {
    "startMs": 531200,
    "text": "earliest case study we did on how to"
  },
  {
    "startMs": 533080,
    "text": "train models to be very robust."
  },
  {
    "startMs": 535120,
    "text": "So,"
  },
  {
    "startMs": 536120,
    "text": "our research result is a model called"
  },
  {
    "startMs": 538280,
    "text": "the Dyna-1, which is a generalist robot"
  },
  {
    "startMs": 540960,
    "text": "foundation model that's fine-tuned to do"
  },
  {
    "startMs": 542839,
    "text": "napkin folding. They can actually"
  },
  {
    "startMs": 544520,
    "text": "achieve 99.4% success rate over a"
  },
  {
    "startMs": 547680,
    "text": "24-hour span. So, here is a time lapse"
  },
  {
    "startMs": 550800,
    "text": "of the model, you know, doing the task."
  },
  {
    "startMs": 552960,
    "text": "And uh you know, it's sped up about a"
  },
  {
    "startMs": 554720,
    "text": "thousand times, so you can see the clock"
  },
  {
    "startMs": 556720,
    "text": "in the back running very quickly to show"
  },
  {
    "startMs": 558680,
    "text": "the progress. And uh my favorite part of"
  },
  {
    "startMs": 561640,
    "text": "the video is when the clock, you know,"
  },
  {
    "startMs": 563520,
    "text": "hits about like right now, right? Like"
  },
  {
    "startMs": 565520,
    "text": "7:00 a.m. in the morning, so the lights,"
  },
  {
    "startMs": 567320,
    "text": "you know, actually come out in the"
  },
  {
    "startMs": 569000,
    "text": "outside. So, the environments are"
  },
  {
    "startMs": 570920,
    "text": "actually, you know, shifting over time"
  },
  {
    "startMs": 572560,
    "text": "due to the lighting as the robot folds"
  },
  {
    "startMs": 574560,
    "text": "the napkin, but the model's robust"
  },
  {
    "startMs": 576520,
    "text": "enough, and it just keeps going."
  },
  {
    "startMs": 578840,
    "text": "So, how do we actually get to a model"
  },
  {
    "startMs": 580320,
    "text": "that can do this?"
  },
  {
    "startMs": 581720,
    "text": "And uh first of all, you know, this uh"
  },
  {
    "startMs": 583920,
    "text": "video we put out is not a one-time"
  },
  {
    "startMs": 586360,
    "text": "occurrence. The model can do this many,"
  },
  {
    "startMs": 588600,
    "text": "many times"
  },
  {
    "startMs": 590040,
    "text": "uh in the office, you know, so these are"
  },
  {
    "startMs": 591960,
    "text": "four distinct trials of 24-hour runs."
  },
  {
    "startMs": 595960,
    "text": "So, before I dive into the technical"
  },
  {
    "startMs": 597480,
    "text": "detail, just to highlight how difficult"
  },
  {
    "startMs": 599280,
    "text": "the task is. So, you see that in the"
  },
  {
    "startMs": 601520,
    "text": "napkin folding task, you start out with"
  },
  {
    "startMs": 603280,
    "text": "a stack of napkin on the side, and what"
  },
  {
    "startMs": 605440,
    "text": "the robot has to do is like be very"
  },
  {
    "startMs": 607120,
    "text": "precise about picking out exactly one"
  },
  {
    "startMs": 609120,
    "text": "napkin from the stack, and then fold it,"
  },
  {
    "startMs": 611360,
    "text": "and then put it into a bin. And then a"
  },
  {
    "startMs": 613240,
    "text": "lot of the"
  },
  {
    "startMs": 614200,
    "text": "failure cases just come from the fact"
  },
  {
    "startMs": 615880,
    "text": "that these, you know, parallel jaw"
  },
  {
    "startMs": 617120,
    "text": "grippers we're deploying, you know, the"
  },
  {
    "startMs": 618960,
    "text": "gripper may not be precise enough to be"
  },
  {
    "startMs": 620720,
    "text": "able to pick out exactly one napkin. And"
  },
  {
    "startMs": 623280,
    "text": "in these situations, the model has to"
  },
  {
    "startMs": 625080,
    "text": "learn how to recover from the mistakes"
  },
  {
    "startMs": 627520,
    "text": "of pulling out extra napkins."
  },
  {
    "startMs": 629880,
    "text": "And then secondly, the in commercial"
  },
  {
    "startMs": 632160,
    "text": "environment, different from a lab demo"
  },
  {
    "startMs": 634440,
    "text": "where the researchers like me are"
  },
  {
    "startMs": 636000,
    "text": "thinking of task success, there's"
  },
  {
    "startMs": 638440,
    "text": "actually very"
  },
  {
    "startMs": 640080,
    "text": "well-defined success criteria for these"
  },
  {
    "startMs": 642440,
    "text": "tasks. So, on the right, you see the"
  },
  {
    "startMs": 644400,
    "text": "difference between what we consider a"
  },
  {
    "startMs": 646000,
    "text": "grade five fold, which is a fold quality"
  },
  {
    "startMs": 648760,
    "text": "that the restaurant would accept versus"
  },
  {
    "startMs": 651040,
    "text": "a grade three, which is something that's"
  },
  {
    "startMs": 653200,
    "text": "below the acceptance criteria. And what"
  },
  {
    "startMs": 655440,
    "text": "you see is a barely like 1-in difference"
  },
  {
    "startMs": 657920,
    "text": "in how, you know, low the you know, the"
  },
  {
    "startMs": 659800,
    "text": "second fold seam of the napkin is."
  },
  {
    "startMs": 663760,
    "text": "So, how do we actually get to a model"
  },
  {
    "startMs": 665200,
    "text": "that works really well? So, our internal"
  },
  {
    "startMs": 667240,
    "text": "attempt was kind of like the"
  },
  {
    "startMs": 669040,
    "text": "pre-training"
  },
  {
    "startMs": 670960,
    "text": "many, many hours of data and then"
  },
  {
    "startMs": 673040,
    "text": "post-training recipe that I told you"
  },
  {
    "startMs": 674880,
    "text": "about in the beginning. And doing this"
  },
  {
    "startMs": 676880,
    "text": "roughly gets you about 80% success rate."
  },
  {
    "startMs": 679280,
    "text": "And what happens is that the model, you"
  },
  {
    "startMs": 681720,
    "text": "know, is doing fine in the beginning,"
  },
  {
    "startMs": 683160,
    "text": "but as soon as it makes a mistake, it'll"
  },
  {
    "startMs": 685960,
    "text": "typically go out of distribution, get"
  },
  {
    "startMs": 687880,
    "text": "stuck, and unable to recover."
  },
  {
    "startMs": 690120,
    "text": "So, we have to do something more than"
  },
  {
    "startMs": 691600,
    "text": "the standard pre-training and"
  },
  {
    "startMs": 692800,
    "text": "post-training idea that's very popular"
  },
  {
    "startMs": 695240,
    "text": "in robotics and also in other fields."
  },
  {
    "startMs": 698480,
    "text": "So what we did is that we developed what"
  },
  {
    "startMs": 700680,
    "text": "we consider"
  },
  {
    "startMs": 702080,
    "text": "reward models for complex long horizon"
  },
  {
    "startMs": 704560,
    "text": "manipulation tasks. So these are models"
  },
  {
    "startMs": 706920,
    "text": "that can, you know, look at a robot"
  },
  {
    "startMs": 708400,
    "text": "video and accurately score its progress"
  },
  {
    "startMs": 711320,
    "text": "towards solving a task. All right, so"
  },
  {
    "startMs": 713440,
    "text": "let me just play these videos again. So"
  },
  {
    "startMs": 715720,
    "text": "here's the same model, you know, being"
  },
  {
    "startMs": 717520,
    "text": "able to score how well the robot is"
  },
  {
    "startMs": 719840,
    "text": "doing these long horizon complex tasks"
  },
  {
    "startMs": 722520,
    "text": "as it's, you know, going from"
  },
  {
    "startMs": 724760,
    "text": "you know, starting of the task to"
  },
  {
    "startMs": 726280,
    "text": "finish. So you see that as the robot's"
  },
  {
    "startMs": 728400,
    "text": "completing a task, it's able to go from"
  },
  {
    "startMs": 730280,
    "text": "zero to one. And if you squint at these"
  },
  {
    "startMs": 732600,
    "text": "videos enough, you also see that"
  },
  {
    "startMs": 734440,
    "text": "whenever the robot is actually making"
  },
  {
    "startMs": 736480,
    "text": "some mistakes, there will be like slight"
  },
  {
    "startMs": 738480,
    "text": "dips in the reward model. And that"
  },
  {
    "startMs": 740880,
    "text": "actually becomes a very important"
  },
  {
    "startMs": 743200,
    "text": "insight into how to make these models"
  },
  {
    "startMs": 745680,
    "text": "very robust. So here's what happens when"
  },
  {
    "startMs": 748040,
    "text": "you run such model during like a"
  },
  {
    "startMs": 750120,
    "text": "autonomous run out of the robot. So you"
  },
  {
    "startMs": 752280,
    "text": "see that when the robot is doing fine"
  },
  {
    "startMs": 754720,
    "text": "folding napkins, the progress estimation"
  },
  {
    "startMs": 757040,
    "text": "is roughly monotonic going up, right?"
  },
  {
    "startMs": 759720,
    "text": "Because the robots are not messing up."
  },
  {
    "startMs": 761600,
    "text": "But what's really interesting is is that"
  },
  {
    "startMs": 763480,
    "text": "let me just for fast forward a bit."
  },
  {
    "startMs": 765640,
    "text": "Whenever the model starts to make"
  },
  {
    "startMs": 767440,
    "text": "mistakes, so here it is. This is the"
  },
  {
    "startMs": 770000,
    "text": "ninth napkin is folding. So you see that"
  },
  {
    "startMs": 772240,
    "text": "when the model is like making mistake,"
  },
  {
    "startMs": 774080,
    "text": "that's when the progress estimation, you"
  },
  {
    "startMs": 776160,
    "text": "know, starts to like, you know, show"
  },
  {
    "startMs": 777720,
    "text": "non-monotonic sign indicating that the"
  },
  {
    "startMs": 779560,
    "text": "robot is messing up. Right? And this is"
  },
  {
    "startMs": 781720,
    "text": "very important because once your model"
  },
  {
    "startMs": 783600,
    "text": "is like good enough in the 90% uh"
  },
  {
    "startMs": 786480,
    "text": "range, then it's very inefficient for"
  },
  {
    "startMs": 788640,
    "text": "humans to manually oversee the robot to"
  },
  {
    "startMs": 791040,
    "text": "detect its failure and then try to"
  },
  {
    "startMs": 792960,
    "text": "recover. But once we have this reward"
  },
  {
    "startMs": 795280,
    "text": "model, we can actually do what I"
  },
  {
    "startMs": 797200,
    "text": "consider uh scalable supervision. So you"
  },
  {
    "startMs": 799920,
    "text": "can just have the robot trying to fold"
  },
  {
    "startMs": 801640,
    "text": "napkins and then run this reward model"
  },
  {
    "startMs": 803680,
    "text": "in the background. So whenever a model"
  },
  {
    "startMs": 805520,
    "text": "does make a mistake, uh we as"
  },
  {
    "startMs": 807840,
    "text": "researchers or operators can immediately"
  },
  {
    "startMs": 810000,
    "text": "know the kind of video case that the"
  },
  {
    "startMs": 811840,
    "text": "model is struggling on and then do very"
  },
  {
    "startMs": 814080,
    "text": "targeted data collection and error"
  },
  {
    "startMs": 816280,
    "text": "recovery data for the model, then we can"
  },
  {
    "startMs": 818920,
    "text": "fine-tune the model again. So, the"
  },
  {
    "startMs": 820880,
    "text": "overall pipeline looks like a"
  },
  {
    "startMs": 822440,
    "text": "human-in-the-loop active learning"
  },
  {
    "startMs": 824080,
    "text": "process where we can use the reward"
  },
  {
    "startMs": 826000,
    "text": "model to help us catch the the kind of"
  },
  {
    "startMs": 828840,
    "text": "mistake the model is bad at and then do"
  },
  {
    "startMs": 831040,
    "text": "targeted collection to make the model"
  },
  {
    "startMs": 832680,
    "text": "better and iterate."
  },
  {
    "startMs": 834440,
    "text": "And what we found is that once you"
  },
  {
    "startMs": 835880,
    "text": "iterate on this"
  },
  {
    "startMs": 837520,
    "text": "uh couple cycles, then you start to get"
  },
  {
    "startMs": 840120,
    "text": "a model that's extremely robust and can"
  },
  {
    "startMs": 842160,
    "text": "recover from all kinds of errors and"
  },
  {
    "startMs": 844520,
    "text": "finally bringing us closer to a"
  },
  {
    "startMs": 846320,
    "text": "commercial grade robots. So, here is"
  },
  {
    "startMs": 848520,
    "text": "just uh some of the"
  },
  {
    "startMs": 850320,
    "text": "uh"
  },
  {
    "startMs": 851160,
    "text": "you know"
  },
  {
    "startMs": 852160,
    "text": "uh"
  },
  {
    "startMs": 853040,
    "text": "highlights, I guess, during the 24-hour"
  },
  {
    "startMs": 855640,
    "text": "trial. So, what you saw there was the"
  },
  {
    "startMs": 857240,
    "text": "robot accidentally picked out more than"
  },
  {
    "startMs": 859400,
    "text": "one napkins and the model is able to,"
  },
  {
    "startMs": 861880,
    "text": "you know, separate the napkins and uh"
  },
  {
    "startMs": 864320,
    "text": "you know, here it's kind of doing that"
  },
  {
    "startMs": 866160,
    "text": "and uh"
  },
  {
    "startMs": 866920,
    "text": "be able to continue progressing. And"
  },
  {
    "startMs": 870080,
    "text": "what we found very interesting is that"
  },
  {
    "startMs": 872200,
    "text": "uh"
  },
  {
    "startMs": 873080,
    "text": "napkin folding is a deformable object"
  },
  {
    "startMs": 874920,
    "text": "manipulation task, right? So, there is"
  },
  {
    "startMs": 876600,
    "text": "almost infinitely many possible states"
  },
  {
    "startMs": 879240,
    "text": "or configuration that a napkin can get"
  },
  {
    "startMs": 881080,
    "text": "to. So, it's impossible to exhaustively"
  },
  {
    "startMs": 883120,
    "text": "collect data for all the error cases."
  },
  {
    "startMs": 885760,
    "text": "But once we had done the active learning"
  },
  {
    "startMs": 887440,
    "text": "many, many times, we saw the model able"
  },
  {
    "startMs": 889720,
    "text": "to generalize to new ways of recovery"
  },
  {
    "startMs": 892040,
    "text": "from the mistakes made and continue to"
  },
  {
    "startMs": 893960,
    "text": "make progress. And that contributed to"
  },
  {
    "startMs": 896160,
    "text": "its ability to be able to uh fold that"
  },
  {
    "startMs": 899240,
    "text": "uh 99.4% success rate. So, here's uh"
  },
  {
    "startMs": 902880,
    "text": "what I found the most impressive bit"
  },
  {
    "startMs": 904720,
    "text": "from the trial. So, typically, if you"
  },
  {
    "startMs": 907320,
    "text": "look at robot videos, you know, they"
  },
  {
    "startMs": 908800,
    "text": "only show you the successful cases. But"
  },
  {
    "startMs": 910960,
    "text": "here, I wanted to highlight that even"
  },
  {
    "startMs": 912920,
    "text": "where a model accidentally pulled the"
  },
  {
    "startMs": 914560,
    "text": "entire napkins stack over, it's able to"
  },
  {
    "startMs": 916800,
    "text": "demonstrate this kind of error recovery"
  },
  {
    "startMs": 918520,
    "text": "behavior. That was very surprising to us"
  },
  {
    "startMs": 920320,
    "text": "when we were developing the model and it"
  },
  {
    "startMs": 922080,
    "text": "contribute to how it's able to just"
  },
  {
    "startMs": 923920,
    "text": "continuously run, right? So, here"
  },
  {
    "startMs": 927000,
    "text": "he made a"
  },
  {
    "startMs": 928120,
    "text": "big mess, but he's able to just do all"
  },
  {
    "startMs": 930480,
    "text": "kind of like very impressive like"
  },
  {
    "startMs": 932600,
    "text": "pulling you know, stretching behavior to"
  },
  {
    "startMs": 935000,
    "text": "recover from his mistake and then"
  },
  {
    "startMs": 936800,
    "text": "continuously going."
  },
  {
    "startMs": 939400,
    "text": "Yeah, so you know, after developing such"
  },
  {
    "startMs": 941760,
    "text": "technology, we actually are successful"
  },
  {
    "startMs": 944240,
    "text": "at deploying our models at many many"
  },
  {
    "startMs": 946720,
    "text": "restaurants in the US. So, here's a real"
  },
  {
    "startMs": 949520,
    "text": "restaurant deployment of our robot"
  },
  {
    "startMs": 951800,
    "text": "folding napkins for the customers in"
  },
  {
    "startMs": 953760,
    "text": "their back office."
  },
  {
    "startMs": 956240,
    "text": "And in addition to folding napkin,"
  },
  {
    "startMs": 958440,
    "text": "because the recipe is quite"
  },
  {
    "startMs": 959640,
    "text": "generalizable, now we also have models"
  },
  {
    "startMs": 961640,
    "text": "doing bunch of commercial tasks. So,"
  },
  {
    "startMs": 963640,
    "text": "here's at a real laundromat in"
  },
  {
    "startMs": 966400,
    "text": "Sacramento, and this time we fine-tune"
  },
  {
    "startMs": 969240,
    "text": "our model to do"
  },
  {
    "startMs": 970880,
    "text": "towel folding for the customer. So, you"
  },
  {
    "startMs": 973160,
    "text": "can see you know, there's a restaurant I"
  },
  {
    "startMs": 975640,
    "text": "guess a laundromat worker coming here to"
  },
  {
    "startMs": 977600,
    "text": "fill the bin over and over, and the"
  },
  {
    "startMs": 979440,
    "text": "robot just kept folding stacks of towels"
  },
  {
    "startMs": 982520,
    "text": "to serve the customer."
  },
  {
    "startMs": 987000,
    "text": "So, now let's talk about we have a"
  },
  {
    "startMs": 989280,
    "text": "recipe to master very complex tasks."
  },
  {
    "startMs": 992160,
    "text": "But, the caveat here is that in all the"
  },
  {
    "startMs": 994000,
    "text": "videos I've shown you so far, we have"
  },
  {
    "startMs": 995640,
    "text": "also collected data at the exact"
  },
  {
    "startMs": 998280,
    "text": "customer site where location that the"
  },
  {
    "startMs": 1000080,
    "text": "robot is deployed. But, if you think"
  },
  {
    "startMs": 1001880,
    "text": "about scaling robots to any task or to"
  },
  {
    "startMs": 1005280,
    "text": "any customer site, then it'll be much"
  },
  {
    "startMs": 1007320,
    "text": "better or more ideal if the models can"
  },
  {
    "startMs": 1009960,
    "text": "readily generalize the environments they"
  },
  {
    "startMs": 1011640,
    "text": "haven't seen before. So, this is what we"
  },
  {
    "startMs": 1014240,
    "text": "have worked on in the"
  },
  {
    "startMs": 1016640,
    "text": "I guess the end of last year where we"
  },
  {
    "startMs": 1018160,
    "text": "figured out a data recipe to collect a"
  },
  {
    "startMs": 1020520,
    "text": "lot of diverse tasks to allow the robots"
  },
  {
    "startMs": 1022880,
    "text": "to be able to deploy at a new site"
  },
  {
    "startMs": 1024720,
    "text": "without any additional data while"
  },
  {
    "startMs": 1026760,
    "text": "maintaining the task performance. So,"
  },
  {
    "startMs": 1029199,
    "text": "here's a demo we did at Coral 2025. So,"
  },
  {
    "startMs": 1032360,
    "text": "Coral is the premier academic conference"
  },
  {
    "startMs": 1034760,
    "text": "on robot learning, and it's held in"
  },
  {
    "startMs": 1036600,
    "text": "Korea last year. So, you know, bringing"
  },
  {
    "startMs": 1039240,
    "text": "a robot from the US to Korea was its"
  },
  {
    "startMs": 1041640,
    "text": "whole challenge that I can talk about"
  },
  {
    "startMs": 1043319,
    "text": "offline. But, the recipe we discovered"
  },
  {
    "startMs": 1046120,
    "text": "was able to, you know, we just brought"
  },
  {
    "startMs": 1048160,
    "text": "the robot to our, you know, exhibition"
  },
  {
    "startMs": 1050320,
    "text": "booth and just dropped it there, and the"
  },
  {
    "startMs": 1052240,
    "text": "robot can start folding many, many"
  },
  {
    "startMs": 1054280,
    "text": "different t-shirts. And you see that the"
  },
  {
    "startMs": 1056680,
    "text": "robot is facing the, you know, the"
  },
  {
    "startMs": 1058400,
    "text": "conference attendees. So, many, many"
  },
  {
    "startMs": 1060800,
    "text": "times there were people just"
  },
  {
    "startMs": 1062080,
    "text": "deliberately trying to mess with the"
  },
  {
    "startMs": 1063440,
    "text": "robot, use the t-shirt to cover up the"
  },
  {
    "startMs": 1065640,
    "text": "robot's camera, then, you know, all kind"
  },
  {
    "startMs": 1067480,
    "text": "of fancy stuff that you see at academic"
  },
  {
    "startMs": 1069200,
    "text": "conferences. But, the model is able to"
  },
  {
    "startMs": 1071240,
    "text": "continuously fold t-shirts over and over"
  },
  {
    "startMs": 1073840,
    "text": "for 3 days straight at the conference"
  },
  {
    "startMs": 1075720,
    "text": "just to demonstrate the ability to, you"
  },
  {
    "startMs": 1078200,
    "text": "know, solve this task at environments"
  },
  {
    "startMs": 1080880,
    "text": "it's never seen before, bringing it, you"
  },
  {
    "startMs": 1082800,
    "text": "know, much closer to the kind of, you"
  },
  {
    "startMs": 1084360,
    "text": "know, ideal, you know, go-to-market that"
  },
  {
    "startMs": 1087480,
    "text": "you want to have for robotics company,"
  },
  {
    "startMs": 1089360,
    "text": "which is just putting the robot to a"
  },
  {
    "startMs": 1090760,
    "text": "site and it just starts working."
  },
  {
    "startMs": 1093680,
    "text": "And recently, we have also ventured into"
  },
  {
    "startMs": 1096080,
    "text": "many different tasks. Like, we have a a"
  },
  {
    "startMs": 1098400,
    "text": "partnership with Red Bull where we're"
  },
  {
    "startMs": 1100560,
    "text": "opening Red Bull cans at, you know, Red"
  },
  {
    "startMs": 1103200,
    "text": "Bull events. So, here's a video of our"
  },
  {
    "startMs": 1108000,
    "text": "uh"
  },
  {
    "startMs": 1108800,
    "text": "robot. Again, uh I guess this video"
  },
  {
    "startMs": 1111000,
    "text": "won't play properly. So, but you get the"
  },
  {
    "startMs": 1113600,
    "text": "point. We brought the robot to a Red"
  },
  {
    "startMs": 1115480,
    "text": "Bull event and it's able to"
  },
  {
    "startMs": 1117200,
    "text": "uh open uh these Red Bull drinks for,"
  },
  {
    "startMs": 1120560,
    "text": "again, conference attendees over and"
  },
  {
    "startMs": 1122280,
    "text": "over without failure, even though in"
  },
  {
    "startMs": 1124280,
    "text": "this particular deployment, it's at a"
  },
  {
    "startMs": 1126640,
    "text": "music festival, so the lighting's always"
  },
  {
    "startMs": 1128560,
    "text": "changing in the background, but the"
  },
  {
    "startMs": 1129880,
    "text": "model can continue. Yeah, but due to"
  },
  {
    "startMs": 1132920,
    "text": "yeah, I guess the"
  },
  {
    "startMs": 1134840,
    "text": "video will not play."
  },
  {
    "startMs": 1136600,
    "text": "So, just to summarize, uh our mission at"
  },
  {
    "startMs": 1139400,
    "text": "AINA is to be able to build"
  },
  {
    "startMs": 1141520,
    "text": "general-purpose models and robots that's"
  },
  {
    "startMs": 1143680,
    "text": "both competent at many tasks, but also"
  },
  {
    "startMs": 1146360,
    "text": "being able to focus and attain really"
  },
  {
    "startMs": 1148640,
    "text": "high performance on commercial tasks."
  },
  {
    "startMs": 1151240,
    "text": "And I have demonstrated some of our"
  },
  {
    "startMs": 1153680,
    "text": "recent progress on mastering complex"
  },
  {
    "startMs": 1155760,
    "text": "tasks and also generalizing the skills"
  },
  {
    "startMs": 1157760,
    "text": "to environments is the robots have never"
  },
  {
    "startMs": 1159800,
    "text": "seen before. So, we are a series A stage"
  },
  {
    "startMs": 1162800,
    "text": "company actively growing and hiring and"
  },
  {
    "startMs": 1164960,
    "text": "if you're also interested in partnering"
  },
  {
    "startMs": 1166480,
    "text": "with us for deployments and other"
  },
  {
    "startMs": 1168160,
    "text": "things, feel free to reach out and then"
  },
  {
    "startMs": 1170080,
    "text": "talk to me after the talk and thank you"
  },
  {
    "startMs": 1172000,
    "text": "for listening."
  },
  {
    "startMs": 1173280,
    "text": ">> Awesome. Thank you, Jason. Really"
  },
  {
    "startMs": 1174800,
    "text": "appreciate."
  },
  {
    "startMs": 1176040,
    "text": "If you guys haven't seen the data robot"
  },
  {
    "startMs": 1177640,
    "text": "in life in real life, have you check it"
  },
  {
    "startMs": 1179920,
    "text": "out? I will say see as very impressive"
  },
  {
    "startMs": 1181600,
    "text": "performance."
  },
  {
    "startMs": 1183680,
    "text": "Do we have any question or want to ask?"
  },
  {
    "startMs": 1186280,
    "text": "Okay, cool."
  },
  {
    "startMs": 1191960,
    "text": ">> Thank you for sharing the presentation."
  },
  {
    "startMs": 1194360,
    "text": "I wanted to ask if you are already doing"
  },
  {
    "startMs": 1197040,
    "text": "kind of like"
  },
  {
    "startMs": 1198440,
    "text": "providing developer kits with robots,"
  },
  {
    "startMs": 1201200,
    "text": "SDKs, frameworks, etc. for your"
  },
  {
    "startMs": 1203400,
    "text": "commercial partners?"
  },
  {
    "startMs": 1205000,
    "text": ">> Yeah, so we haven't been working on"
  },
  {
    "startMs": 1207480,
    "text": "develop developer kit, but you know,"
  },
  {
    "startMs": 1210840,
    "text": "at the current moment we're building our"
  },
  {
    "startMs": 1212760,
    "text": "own hardware stack. So, maybe at some"
  },
  {
    "startMs": 1215360,
    "text": "point it's on our road map, but"
  },
  {
    "startMs": 1217560,
    "text": "not as of not as of now. Yeah."
  },
  {
    "startMs": 1226320,
    "text": ">> I'm curious if you have ventured into"
  },
  {
    "startMs": 1228800,
    "text": "education like teaching kids"
  },
  {
    "startMs": 1232640,
    "text": "elementary maths or any such use case"
  },
  {
    "startMs": 1235240,
    "text": "you tried."
  },
  {
    "startMs": 1236120,
    "text": ">> Yeah, so we haven't looked into"
  },
  {
    "startMs": 1237880,
    "text": "education use case, right?"
  },
  {
    "startMs": 1240400,
    "text": "But I think in the future, you know,"
  },
  {
    "startMs": 1241600,
    "text": "once our"
  },
  {
    "startMs": 1243080,
    "text": "full stack robotics pipelines mature,"
  },
  {
    "startMs": 1245040,
    "text": "once we have our own hardware, our data"
  },
  {
    "startMs": 1246640,
    "text": "collection, you know, toolkit, I think"
  },
  {
    "startMs": 1249120,
    "text": "it's possible and I think it'll be very"
  },
  {
    "startMs": 1250840,
    "text": "interesting to venture into education"
  },
  {
    "startMs": 1253280,
    "text": "use cases because I also think robotics"
  },
  {
    "startMs": 1255880,
    "text": "will only get bigger in the future. So,"
  },
  {
    "startMs": 1257960,
    "text": "I think it'll be really really"
  },
  {
    "startMs": 1258960,
    "text": "interesting to get children and young"
  },
  {
    "startMs": 1261120,
    "text": "kids into the field and the learning how"
  },
  {
    "startMs": 1263160,
    "text": "to actually train models to do tasks."
  },
  {
    "startMs": 1267240,
    "text": ">> Thank you for the presentation, really"
  },
  {
    "startMs": 1268720,
    "text": "good. Uh"
  },
  {
    "startMs": 1269760,
    "text": "my question is like is there do you guys"
  },
  {
    "startMs": 1271560,
    "text": "have a plan for uh kind of taking this"
  },
  {
    "startMs": 1274240,
    "text": "technology direct to consumer? Like I"
  },
  {
    "startMs": 1276760,
    "text": "saw the use case for the laundry folding"
  },
  {
    "startMs": 1279400,
    "text": "laundry."
  },
  {
    "startMs": 1280560,
    "text": "A lot of people don't like folding"
  },
  {
    "startMs": 1281880,
    "text": "laundry. I think this is a great use"
  },
  {
    "startMs": 1283320,
    "text": "case. So, is that something that you"
  },
  {
    "startMs": 1285880,
    "text": "guys thinking about? And obviously"
  },
  {
    "startMs": 1287720,
    "text": "there's a cost and all those aspects,"
  },
  {
    "startMs": 1289680,
    "text": "but what's what's your plan in the long"
  },
  {
    "startMs": 1291920,
    "text": "term on this technology?"
  },
  {
    "startMs": 1293400,
    "text": ">> Yeah, so our long-term plan is to"
  },
  {
    "startMs": 1295000,
    "text": "develop, you know, like a robot model"
  },
  {
    "startMs": 1297320,
    "text": "plus hardware platform that can be"
  },
  {
    "startMs": 1299000,
    "text": "deployed anywhere for anyone. So, that"
  },
  {
    "startMs": 1300720,
    "text": "would include, you know, like going"
  },
  {
    "startMs": 1302520,
    "text": "directly to consumers. And you know, our"
  },
  {
    "startMs": 1304720,
    "text": "models today are able to just fold, you"
  },
  {
    "startMs": 1306720,
    "text": "know, uh many kinds of garments. But uh"
  },
  {
    "startMs": 1309720,
    "text": "in terms of like go to market strategy,"
  },
  {
    "startMs": 1311520,
    "text": "our current focus is on enterprise use"
  },
  {
    "startMs": 1313560,
    "text": "cases because I think the distribution"
  },
  {
    "startMs": 1315760,
    "text": "channel is like I think it's easier to"
  },
  {
    "startMs": 1318480,
    "text": "uh work with uh"
  },
  {
    "startMs": 1320200,
    "text": "enterprise customers. Like we can deploy"
  },
  {
    "startMs": 1322280,
    "text": "right away and iterate at customer"
  },
  {
    "startMs": 1324120,
    "text": "sites. But for consumer product, I think"
  },
  {
    "startMs": 1327200,
    "text": "uh I would imagine that it has to be"
  },
  {
    "startMs": 1329360,
    "text": "very very polished and it's less"
  },
  {
    "startMs": 1330920,
    "text": "tolerant for mistakes. And there is a"
  },
  {
    "startMs": 1332800,
    "text": "lot of privacy safety concerns that we"
  },
  {
    "startMs": 1334840,
    "text": "think uh we are"
  },
  {
    "startMs": 1337000,
    "text": "trying to not get into at the current"
  },
  {
    "startMs": 1339120,
    "text": "moment to"
  },
  {
    "startMs": 1340880,
    "text": "uh unblock us from deployments because"
  },
  {
    "startMs": 1343160,
    "text": "we think that the bottleneck for AI"
  },
  {
    "startMs": 1345280,
    "text": "robot is getting AI and hardware"
  },
  {
    "startMs": 1348200,
    "text": "co-working together very very well. And"
  },
  {
    "startMs": 1349960,
    "text": "I think commercial environments provide"
  },
  {
    "startMs": 1351520,
    "text": "ideal testing ground uh for companies"
  },
  {
    "startMs": 1354440,
    "text": "and for the entire field at this stage."
  },
  {
    "startMs": 1358480,
    "text": ">> Thank you, Jason."
  },
  {
    "startMs": 1360240,
    "text": "Uh my name is Ahmed and I work in a"
  },
  {
    "startMs": 1363440,
    "text": "in voice AI in in the voice space. And"
  },
  {
    "startMs": 1366360,
    "text": "at least for consumer robots, uh"
  },
  {
    "startMs": 1369320,
    "text": "where I believe humans will want to"
  },
  {
    "startMs": 1371760,
    "text": "issue voice commands to robots, uh can"
  },
  {
    "startMs": 1374360,
    "text": "you talk about how to integrate uh"
  },
  {
    "startMs": 1377160,
    "text": "traditional voice stacks, you know,"
  },
  {
    "startMs": 1379280,
    "text": "which are intent-driven and uh uh"
  },
  {
    "startMs": 1381680,
    "text": "different kinds of models from uh uh"
  },
  {
    "startMs": 1384200,
    "text": "robotic, you know, VLA models."
  },
  {
    "startMs": 1386640,
    "text": ">> Yeah, that's a great question. So, we"
  },
  {
    "startMs": 1388240,
    "text": "think uh robot human interactions really"
  },
  {
    "startMs": 1391120,
    "text": "important, and it at the end of the day,"
  },
  {
    "startMs": 1393000,
    "text": "right? We want to develop models that"
  },
  {
    "startMs": 1394520,
    "text": "are teachable. So, it'd be really nice"
  },
  {
    "startMs": 1396360,
    "text": "if humans can speak to it, and the robot"
  },
  {
    "startMs": 1398360,
    "text": "does the task. So, the way I think about"
  },
  {
    "startMs": 1400440,
    "text": "it, so I'm not I don't have a speech or"
  },
  {
    "startMs": 1402160,
    "text": "audio background, so what I can think of"
  },
  {
    "startMs": 1403880,
    "text": "is like there is very mature"
  },
  {
    "startMs": 1405920,
    "text": "speech-to-text models, right? So, you"
  },
  {
    "startMs": 1407800,
    "text": "can use that running the on-board to"
  },
  {
    "startMs": 1410880,
    "text": "translate human intent or human speech"
  },
  {
    "startMs": 1413720,
    "text": "into text very quickly, and then, you"
  },
  {
    "startMs": 1416360,
    "text": "know, that's where our reasoning model"
  },
  {
    "startMs": 1418200,
    "text": "and our world model comes into play,"
  },
  {
    "startMs": 1419920,
    "text": "because both of them are able to"
  },
  {
    "startMs": 1421640,
    "text": "interpret text commands. So, that will"
  },
  {
    "startMs": 1423800,
    "text": "allow the model to translate instruction"
  },
  {
    "startMs": 1426880,
    "text": "into actions. But, I think in the"
  },
  {
    "startMs": 1429600,
    "text": "overall stack, the bottleneck is still"
  },
  {
    "startMs": 1431640,
    "text": "on the robot foundation models, because"
  },
  {
    "startMs": 1433880,
    "text": "models today our models and other"
  },
  {
    "startMs": 1436080,
    "text": "people's model are not ready to just"
  },
  {
    "startMs": 1437960,
    "text": "execute any arbitrary commands. So, I"
  },
  {
    "startMs": 1439720,
    "text": "think there's still a gap in terms of"
  },
  {
    "startMs": 1441600,
    "text": "low-level robotic manipulation to get"
  },
  {
    "startMs": 1443520,
    "text": "there. But, I think that's the eventual"
  },
  {
    "startMs": 1445600,
    "text": "goal that we want to get to. Just a"
  },
  {
    "startMs": 1447080,
    "text": "model that can be steerable, that can be"
  },
  {
    "startMs": 1449240,
    "text": "taught by anyone. Yeah."
  },
  {
    "startMs": 1452800,
    "text": "Okay. Thank you."
  },
  {
    "startMs": 1454040,
    "text": ">> Um how do you think about, I guess, the"
  },
  {
    "startMs": 1456600,
    "text": "split of intelligence between"
  },
  {
    "startMs": 1459640,
    "text": "the general foundation models that you"
  },
  {
    "startMs": 1461200,
    "text": "all are training, I guess, something"
  },
  {
    "startMs": 1463200,
    "text": "like a task-specific model for something"
  },
  {
    "startMs": 1464880,
    "text": "like laundry folding, and then kind of"
  },
  {
    "startMs": 1466680,
    "text": "some of the on-site specific training"
  },
  {
    "startMs": 1469160,
    "text": "that needs to happen for a specific"
  },
  {
    "startMs": 1470800,
    "text": "deployment in a particular environment?"
  },
  {
    "startMs": 1473960,
    "text": ">> Yeah, I think both are very very"
  },
  {
    "startMs": 1476040,
    "text": "important, right? So, the kind of recipe"
  },
  {
    "startMs": 1478720,
    "text": "that we have is, you know, we have"
  },
  {
    "startMs": 1479880,
    "text": "pre-training, post-training, then we"
  },
  {
    "startMs": 1481280,
    "text": "have some of the sort of active"
  },
  {
    "startMs": 1483320,
    "text": "learning, right? I I think"
  },
  {
    "startMs": 1485200,
    "text": "uh"
  },
  {
    "startMs": 1485760,
    "text": "if you think about, you know, humans,"
  },
  {
    "startMs": 1487640,
    "text": "right? You know, we have like basic"
  },
  {
    "startMs": 1489400,
    "text": "level of like semantic and physical"
  },
  {
    "startMs": 1491320,
    "text": "understanding that allows us to adapt"
  },
  {
    "startMs": 1493120,
    "text": "any physical task very fast, and I think"
  },
  {
    "startMs": 1495560,
    "text": "the best way to even build commercial"
  },
  {
    "startMs": 1497920,
    "text": "grade robots today is by doing that."
  },
  {
    "startMs": 1501240,
    "text": "Because by doing a lot of generalist"
  },
  {
    "startMs": 1503360,
    "text": "model training have a general"
  },
  {
    "startMs": 1504640,
    "text": "understanding of physical world that's"
  },
  {
    "startMs": 1506640,
    "text": "very useful. And what we have seen, so"
  },
  {
    "startMs": 1508880,
    "text": "this is something I didn't talk about in"
  },
  {
    "startMs": 1510160,
    "text": "the talk is that when the model was"
  },
  {
    "startMs": 1512640,
    "text": "recovering from all kind of errors that"
  },
  {
    "startMs": 1514520,
    "text": "you saw, a lot of that also just came"
  },
  {
    "startMs": 1516560,
    "text": "from interpolating different error"
  },
  {
    "startMs": 1518080,
    "text": "recovery behavior that is gathered from"
  },
  {
    "startMs": 1520480,
    "text": "doing other tasks, right? Just because,"
  },
  {
    "startMs": 1522640,
    "text": "you know, the model has seen thousands"
  },
  {
    "startMs": 1523960,
    "text": "of tasks we had some data to recover"
  },
  {
    "startMs": 1526760,
    "text": "from all kinds of mistakes is able to"
  },
  {
    "startMs": 1529000,
    "text": "just execute that kind of like uh on"
  },
  {
    "startMs": 1532040,
    "text": "demand on a new task it hasn't seen"
  },
  {
    "startMs": 1533760,
    "text": "before. But if you're only specializing"
  },
  {
    "startMs": 1535400,
    "text": "a model from the get-go on one task,"
  },
  {
    "startMs": 1537560,
    "text": "then it's actually missing out on a lot"
  },
  {
    "startMs": 1539440,
    "text": "of the physical knowledge that allows"
  },
  {
    "startMs": 1540960,
    "text": "the model to be more robust than if you"
  },
  {
    "startMs": 1543000,
    "text": "only train on one task. And that's what"
  },
  {
    "startMs": 1544640,
    "text": "we have seen consistently. That's why,"
  },
  {
    "startMs": 1547000,
    "text": "you know, in the beginning of the talk I"
  },
  {
    "startMs": 1548240,
    "text": "also emphasize on having that general"
  },
  {
    "startMs": 1550520,
    "text": "pre-training backbone and then doing"
  },
  {
    "startMs": 1552320,
    "text": "post-training and active learning on top"
  },
  {
    "startMs": 1554000,
    "text": "of that to get to actual commercial"
  },
  {
    "startMs": 1555760,
    "text": "grade usability, not on just one task,"
  },
  {
    "startMs": 1558520,
    "text": "but also the same recipe that can be"
  },
  {
    "startMs": 1560400,
    "text": "repeated across many many tasks."
  },
  {
    "startMs": 1563320,
    "text": ">> Awesome. Thank you, Jason. Really"
  },
  {
    "startMs": 1564640,
    "text": "appreciate. Um that's a wrap for uh data"
  },
  {
    "startMs": 1567520,
    "text": "session. Uh if you have more questions,"
  },
  {
    "startMs": 1569680,
    "text": "feel free to catch Jason after the"
  },
  {
    "startMs": 1570880,
    "text": "session as well. Uh but yeah, this wrap"
  },
  {
    "startMs": 1573080,
    "text": "our morning session. Uh this afternoon"
  },
  {
    "startMs": 1575120,
    "text": "we have Unity, Skydio, Zoox, uh Waymo,"
  },
  {
    "startMs": 1577680,
    "text": "DeepMind. Um so we'll catch you and see"
  },
  {
    "startMs": 1579800,
    "text": "you AI as well. So we'll see you in the"
  },
  {
    "startMs": 1581440,
    "text": "afternoon. Thank you, everyone."
  },
  {
    "startMs": 1599528,
    "text": ">> [music]"
  }
]