[
  {
    "start": 38.204,
    "end": 39.164,
    "text": "How's it going, everyone?"
  },
  {
    "start": 39.164,
    "end": 40.165,
    "text": "My name is Logan Kilpatrick."
  },
  {
    "start": 40.165,
    "end": 41.666,
    "text": "I'm part of the Google DeepMind team."
  },
  {
    "start": 41.666,
    "end": 42.876,
    "text": "Welcome back to Release Notes."
  },
  {
    "start": 42.876,
    "end": 43.626,
    "text": "Today we're here"
  },
  {
    "start": 43.626,
    "end": 45.378,
    "text": "in the Gemini Robotics lab,"
  },
  {
    "start": 45.378,
    "end": 48.506,
    "text": "with Carolina, Stuart,\nKanishka, and Jie."
  },
  {
    "start": 48.506,
    "end": 52.052,
    "text": "We're talking about the new\nembodied reasoning model and actually"
  },
  {
    "start": 52.052,
    "end": 56.014,
    "text": "the whole suite of Gemini\nRobotics launches that are coming out."
  },
  {
    "start": 56.014,
    "end": 58.058,
    "text": "So I'm super excited.\nI have a million questions."
  },
  {
    "start": 58.058,
    "end": 59.267,
    "text": "We were talking off camera."
  },
  {
    "start": 59.267,
    "end": 61.227,
    "text": "But Carolina, maybe you can kick us off"
  },
  {
    "start": 61.227,
    "end": 64.355,
    "text": "with sort of the headline\nfor this moment."
  },
  {
    "start": 64.355,
    "end": 66.649,
    "text": "And then we can actually\ntake a step back after that and"
  },
  {
    "start": 66.649,
    "end": 71.321,
    "text": "talk about the arc, maybe,\nacross all the work that it's taken to get here."
  },
  {
    "start": 71.321,
    "end": 74.282,
    "text": "And we'll get to see cool robots,\nwhich I'm excited about."
  },
  {
    "start": 74.282,
    "end": 75.492,
    "text": "- Yeah, definitely."
  },
  {
    "start": 75.492,
    "end": 78.828,
    "text": "What we're building here is\nthe intelligence layer to power"
  },
  {
    "start": 78.828,
    "end": 82.332,
    "text": "any robot to do a\nbroad range of useful tasks."
  },
  {
    "start": 82.332,
    "end": 85.085,
    "text": "And we've been building towards this\nmoment for a very long time, actually."
  },
  {
    "start": 85.085,
    "end": 86.961,
    "text": "Our team has been working"
  },
  {
    "start": 86.961,
    "end": 88.254,
    "text": "towards general-purpose robotics"
  },
  {
    "start": 88.254,
    "end": 90.006,
    "text": "since the inception of the team."
  },
  {
    "start": 90.006,
    "end": 92.801,
    "text": "And we had a long\nhistory of always thinking,"
  },
  {
    "start": 92.801,
    "end": 94.552,
    "text": "how do we solve the problem"
  },
  {
    "start": 94.552,
    "end": 96.596,
    "text": "from first principles in a way"
  },
  {
    "start": 96.596,
    "end": 99.349,
    "text": "that doesn’t just take shortcuts"
  },
  {
    "start": 99.349,
    "end": 101.184,
    "text": "and tries to solve the entire problem?"
  },
  {
    "start": 101.184,
    "end": 102.56,
    "text": "And so the entire problem"
  },
  {
    "start": 102.56,
    "end": 103.645,
    "text": "really means infusing"
  },
  {
    "start": 103.645,
    "end": 106.439,
    "text": "a robot with human-level\nintelligence, right?"
  },
  {
    "start": 106.439,
    "end": 110.401,
    "text": "So that it can understand the\nenvironment as you and I can,"
  },
  {
    "start": 110.401,
    "end": 114.114,
    "text": "so that it can reason about\nwhat it means to complete a task."
  },
  {
    "start": 114.114,
    "end": 116.574,
    "text": "And then it can\nactually take action with all"
  },
  {
    "start": 116.574,
    "end": 119.786,
    "text": "the dexterity that us\nhumans have and take for granted."
  },
  {
    "start": 119.786,
    "end": 121.621,
    "text": "-Yeah.\n-So that's the general goal."
  },
  {
    "start": 121.621,
    "end": 122.831,
    "text": "- I love that."
  },
  {
    "start": 122.831,
    "end": 126.835,
    "text": "And the actual release\nthat we're talking about today"
  },
  {
    "start": 126.835,
    "end": 131.589,
    "text": "is the new package of\nGemini Robotics 2 models,"
  },
  {
    "start": 131.589,
    "end": 135.051,
    "text": "starting with the ER model,\nbut a bunch of other stuff as well."
  },
  {
    "start": 135.051,
    "end": 137.679,
    "text": "Do you want to talk\nthrough the different models,"
  },
  {
    "start": 137.679,
    "end": 139.681,
    "text": "the suite of models\nthat are becoming available?"
  },
  {
    "start": 139.681,
    "end": 142.183,
    "text": "- Yeah, so Gemini Robotics 2 essentially"
  },
  {
    "start": 142.183,
    "end": 144.853,
    "text": "is bringing whole-body\nintelligence to robots."
  },
  {
    "start": 144.853,
    "end": 148.022,
    "text": "So what that means\nis that we are enabling"
  },
  {
    "start": 148.022,
    "end": 152.152,
    "text": "a model that can understand where the entire,\nevery part of the robot"
  },
  {
    "start": 152.152,
    "end": 155.238,
    "text": "is in space, and it can reason\nabout doing more complicated tasks."
  },
  {
    "start": 155.238,
    "end": 157.657,
    "text": "So, imagine that you get\nin your closet and you're"
  },
  {
    "start": 157.657,
    "end": 161.536,
    "text": "trying to clean it up and put\naway all your clothes and your shoes."
  },
  {
    "start": 161.536,
    "end": 165.331,
    "text": "That requires you to move your body\nin all kinds of ways, avoid obstacles,"
  },
  {
    "start": 165.331,
    "end": 167.542,
    "text": "reach things high,\npick things from the floor,"
  },
  {
    "start": 167.542,
    "end": 170.712,
    "text": "this is something that we\ncould not really do before"
  },
  {
    "start": 170.712,
    "end": 173.715,
    "text": "in a way that\nunderstands what's going on."
  },
  {
    "start": 173.715,
    "end": 176.217,
    "text": "We literally put the\nrobot in a garage and ask it,"
  },
  {
    "start": 176.217,
    "end": 177.76,
    "text": "\"Can you please clean up this garage?\""
  },
  {
    "start": 177.76,
    "end": 178.928,
    "text": "Right? And then—"
  },
  {
    "start": 178.928,
    "end": 180.972,
    "text": "- Is that the real prompt\nin those examples?"
  },
  {
    "start": 180.972,
    "end": 183.266,
    "text": "It's literally \"Please clean my garage?\""
  },
  {
    "start": 183.266,
    "end": 187.437,
    "text": "- Yeah, you'll get to see it.\nInspired by our own needs, I think."
  },
  {
    "start": 187.437,
    "end": 188.438,
    "text": "- I love it."
  },
  {
    "start": 188.438,
    "end": 189.939,
    "text": "- So, then the robot\nneeds to reason about what"
  },
  {
    "start": 189.939,
    "end": 191.482,
    "text": "does it mean to clean up a garage?"
  },
  {
    "start": 191.482,
    "end": 192.483,
    "text": "It needs to think about,"
  },
  {
    "start": 192.483,
    "end": 195.486,
    "text": "\"Oh, I'm going to put all the\ncleaning supplies in the same place.\""
  },
  {
    "start": 195.486,
    "end": 196.905,
    "text": "It needs to be able to"
  },
  {
    "start": 196.905,
    "end": 198.448,
    "text": "put things high, up high."
  },
  {
    "start": 198.448,
    "end": 200.992,
    "text": "If something falls on the ground,\nit has to understand that,"
  },
  {
    "start": 200.992,
    "end": 203.203,
    "text": "and then go pick it up,\nand then put it away."
  },
  {
    "start": 203.203,
    "end": 206.206,
    "text": "So that's one of the big things\nthat we're bringing in this release."
  },
  {
    "start": 206.206,
    "end": 209.626,
    "text": "The second thing that we\nworked significantly on,"
  },
  {
    "start": 209.626,
    "end": 212.754,
    "text": "and we still have a lot more work to do,\nis dexterity."
  },
  {
    "start": 212.754,
    "end": 214.756,
    "text": "So, again, us humans take for granted"
  },
  {
    "start": 214.756,
    "end": 217.091,
    "text": "this really dexterous hands\nthat we have,"
  },
  {
    "start": 217.091,
    "end": 220.22,
    "text": "but pretty much everything you\ndo every day requires dexterity."
  },
  {
    "start": 220.22,
    "end": 224.557,
    "text": "And that means folding things,\nopening doors,"
  },
  {
    "start": 224.557,
    "end": 228.061,
    "text": "just picking anything that\nwill fall out of your hands."
  },
  {
    "start": 228.061,
    "end": 231.189,
    "text": "And so, dexterity is an\narea that we worked on a lot."
  },
  {
    "start": 231.189,
    "end": 235.36,
    "text": "And if you just, again,\nif you just get yourself, in the morning, coffee,"
  },
  {
    "start": 235.36,
    "end": 237.07,
    "text": "that is actually a\npretty dexterous task."
  },
  {
    "start": 237.07,
    "end": 239.781,
    "text": "- Yeah.\n- And one thing to know is that"
  },
  {
    "start": 239.781,
    "end": 242.2,
    "text": "because we're building this\nintelligence layer across robots,"
  },
  {
    "start": 242.2,
    "end": 245.536,
    "text": "is we definitely work a lot\non making this work really well"
  },
  {
    "start": 245.536,
    "end": 249.082,
    "text": "on humanoids, but we also bring\nother robots that we're controlling."
  },
  {
    "start": 249.082,
    "end": 252.126,
    "text": "So, the one that is\nright behind you is actually"
  },
  {
    "start": 252.126,
    "end": 255.255,
    "text": "our Franka Duo and\nit has two Franka arms."
  },
  {
    "start": 255.255,
    "end": 257.715,
    "text": "And we also control\nwith the exact same model,"
  },
  {
    "start": 257.715,
    "end": 260.927,
    "text": "this robot in order to do pretty\ndexterous things with those grippers,"
  },
  {
    "start": 260.927,
    "end": 262.637,
    "text": "like pack things neatly."
  },
  {
    "start": 262.637,
    "end": 264.889,
    "text": "So, that's another form of dexterity."
  },
  {
    "start": 264.889,
    "end": 267.517,
    "text": "And then the third thing\nthat we bring in is what we call"
  },
  {
    "start": 267.517,
    "end": 269.185,
    "text": "multi-robot collaboration,"
  },
  {
    "start": 269.185,
    "end": 272.563,
    "text": "which is different robots have\ndifferent capabilities, right?"
  },
  {
    "start": 272.563,
    "end": 275.525,
    "text": "And what we're doing here is that you have,\nyou're bringing the robot"
  },
  {
    "start": 275.525,
    "end": 278.486,
    "text": "the intelligence to know what\nit needs to do to complete a task,"
  },
  {
    "start": 278.486,
    "end": 281.072,
    "text": "but also the\nunderstanding that it can call"
  },
  {
    "start": 281.072,
    "end": 283.7,
    "text": "other robots to accelerate the task,"
  },
  {
    "start": 283.7,
    "end": 287.37,
    "text": "or to do things in parallel,\nand do things faster."
  },
  {
    "start": 287.37,
    "end": 290.331,
    "text": "So, those are roughly the three\nthings that we're doing in this release."
  },
  {
    "start": 290.331,
    "end": 293.293,
    "text": "- I'm excited for this.\nI have a home robot vacuum,"
  },
  {
    "start": 293.293,
    "end": 295.628,
    "text": "and then I have another\nrobot that I have ordered."
  },
  {
    "start": 295.628,
    "end": 298.464,
    "text": "Hopefully it'll be\npowered by Gemini Robotics 2,"
  },
  {
    "start": 298.464,
    "end": 300.55,
    "text": "that doesn’t vacuum but\ndoes a bunch of other stuff."
  },
  {
    "start": 300.55,
    "end": 302.135,
    "text": "And so I feel like this\nis actually going to be—"
  },
  {
    "start": 302.135,
    "end": 304.595,
    "text": "I did not think about this today,\nbut I was thinking to myself,"
  },
  {
    "start": 304.595,
    "end": 307.598,
    "text": "I was like, can I get the one to\nsort of carry the other one around,"
  },
  {
    "start": 307.598,
    "end": 311.519,
    "text": "and go deploy the robot\nto go do certain tasks?"
  },
  {
    "start": 311.519,
    "end": 314.272,
    "text": "But ideally they could just\ncommunicate as robots together"
  },
  {
    "start": 314.272,
    "end": 318.318,
    "text": "and get the work done,\nwhich is super interesting."
  },
  {
    "start": 318.318,
    "end": 322.53,
    "text": "Maybe we can also talk about the arc to get here,\nand obviously,"
  },
  {
    "start": 322.53,
    "end": 324.532,
    "text": "this is the second\nGemini Robotics model,"
  },
  {
    "start": 324.532,
    "end": 326.993,
    "text": "but is there other\nthings that are worth--"
  },
  {
    "start": 326.993,
    "end": 329.287,
    "text": "Obviously, Google's been doing\na lot of robotic stuff."
  },
  {
    "start": 329.287,
    "end": 331.039,
    "text": "DeepMind, maybe, I actually don't know."
  },
  {
    "start": 331.039,
    "end": 334.375,
    "text": "Maybe also had a bunch of robotic stuff,\nbut you can talk about any of"
  },
  {
    "start": 334.375,
    "end": 339.339,
    "text": "the research and the arc\nto get to this launch moment?"
  },
  {
    "start": 339.339,
    "end": 340.965,
    "text": "- Yeah, I mean, it's been a long path."
  },
  {
    "start": 340.965,
    "end": 342.467,
    "text": "I think many of us have been here"
  },
  {
    "start": 342.467,
    "end": 345.72,
    "text": "also for a while and have\nseen all of these steps through."
  },
  {
    "start": 345.72,
    "end": 347.764,
    "text": "But yeah, I mean, we've always,\nfrom the beginning, like I said,"
  },
  {
    "start": 347.764,
    "end": 350.058,
    "text": "we were really thinking\nabout general-purpose robotics"
  },
  {
    "start": 350.058,
    "end": 353.269,
    "text": "before, I think,\nthe field really realized that was feasible."
  },
  {
    "start": 353.269,
    "end": 358.483,
    "text": "And so we've done a few\niterations where we've brought"
  },
  {
    "start": 358.483,
    "end": 360.985,
    "text": "techniques that now are\ntable stakes for the community."
  },
  {
    "start": 360.985,
    "end": 365.865,
    "text": "So for example,\nwe first introduced reinforcement learning,"
  },
  {
    "start": 365.865,
    "end": 370.036,
    "text": "to learning simulation,\nhow to control whole robots, right?"
  },
  {
    "start": 370.036,
    "end": 373.706,
    "text": "So you see a lot of robots today,\ndancing and doing acrobatics."
  },
  {
    "start": 373.706,
    "end": 378.044,
    "text": "Those robots are actually using those\ntechniques in order to move the robots"
  },
  {
    "start": 378.044,
    "end": 382.298,
    "text": "in a way that feels stable and\nthat can mimic a particular sequence."
  },
  {
    "start": 382.298,
    "end": 385.343,
    "text": "We've also shown what's\npossible when it comes to bringing"
  },
  {
    "start": 385.343,
    "end": 388.596,
    "text": "LLMs and VLMs,\nwhen it comes to planning for robots."
  },
  {
    "start": 388.596,
    "end": 392.725,
    "text": "Before that effort, basically,\nrobots did not understand semantics."
  },
  {
    "start": 392.725,
    "end": 394.31,
    "text": "They did not understand our world."
  },
  {
    "start": 394.31,
    "end": 395.937,
    "text": "They did not know what\nyou meant when you say,"
  },
  {
    "start": 395.937,
    "end": 397.063,
    "text": "\"Bring me a cup.\"\n- Yeah."
  },
  {
    "start": 397.063,
    "end": 398.481,
    "text": "- Right? You actually had to say, \"No,"
  },
  {
    "start": 398.481,
    "end": 402.443,
    "text": "bring me that object at position X Y Z,\nin space.\""
  },
  {
    "start": 402.443,
    "end": 404.695,
    "text": "So, that was one clear breakthrough."
  },
  {
    "start": 404.695,
    "end": 408.157,
    "text": "We also introduced transformers to robotics,\nand that shifted the field"
  },
  {
    "start": 408.157,
    "end": 412.453,
    "text": "into this era of data-driven robotics,\nwhere you now had to just"
  },
  {
    "start": 412.453,
    "end": 416.916,
    "text": "collect a lot of data, and teach the\nrobot how to do many different tasks."
  },
  {
    "start": 416.916,
    "end": 419.46,
    "text": "And then introduce\neven the concept of a VLA,"
  },
  {
    "start": 419.46,
    "end": 421.796,
    "text": "which is a new type of\nfoundation model called"
  },
  {
    "start": 421.796,
    "end": 423.548,
    "text": "Vision Language Action (VLA) model,"
  },
  {
    "start": 423.548,
    "end": 426.342,
    "text": "that essentially enables\nrobots to understand natural language"
  },
  {
    "start": 426.342,
    "end": 429.762,
    "text": "and visual input and then directly\ncontrol them in a way that is general."
  },
  {
    "start": 429.762,
    "end": 433.808,
    "text": "So, this VLA type of foundation model"
  },
  {
    "start": 433.808,
    "end": 435.977,
    "text": "is again, adopted in the community."
  },
  {
    "start": 435.977,
    "end": 438.062,
    "text": "And then we've also shown what's\npossible when it comes to dexterity."
  },
  {
    "start": 438.062,
    "end": 439.981,
    "text": "I think many of us\ndid not really think that it"
  },
  {
    "start": 439.981,
    "end": 443.818,
    "text": "was possible to tie shoelaces,\nfor example, in our careers."
  },
  {
    "start": 443.818,
    "end": 446.154,
    "text": "And that's something that,\ntoday, people tie shoelaces,"
  },
  {
    "start": 446.154,
    "end": 448.156,
    "text": "they fold their laundry."
  },
  {
    "start": 448.156,
    "end": 449.907,
    "text": "These are things that we\nshowed that was possible."
  },
  {
    "start": 449.907,
    "end": 453.578,
    "text": "And I think all of\nthat has come together"
  },
  {
    "start": 453.578,
    "end": 456.873,
    "text": "towards this Gemini Robotics\nmodels that we introduced last year."
  },
  {
    "start": 456.873,
    "end": 460.793,
    "text": "And so last year, what we did was we\nbrought all of the power of Gemini's"
  },
  {
    "start": 460.793,
    "end": 464.088,
    "text": "multimodal world understanding,\ncombined with those techniques"
  },
  {
    "start": 464.088,
    "end": 468.509,
    "text": "in order to bring essentially\nGemini's intelligence to robots."
  },
  {
    "start": 468.509,
    "end": 470.428,
    "text": "And all it is, is that we"
  },
  {
    "start": 470.428,
    "end": 474.098,
    "text": "enable Gemini to also\nthink about moving robots."
  },
  {
    "start": 474.098,
    "end": 476.601,
    "text": "So, we add actions as\na modality in Gemini."
  },
  {
    "start": 476.601,
    "end": 481.481,
    "text": "And so that essentially enables\nGemini to understand when you ask it to"
  },
  {
    "start": 481.481,
    "end": 486.694,
    "text": "turn around this, it understands\nwhat it means to move around the bottle."
  },
  {
    "start": 486.694,
    "end": 488.613,
    "text": "Or if you ask it to do\nsomething more complex,"
  },
  {
    "start": 488.613,
    "end": 490.615,
    "text": "pick up all the things that are pink,"
  },
  {
    "start": 490.615,
    "end": 492.325,
    "text": "then it understands what that means."
  },
  {
    "start": 492.325,
    "end": 494.327,
    "text": "So, that's essentially\nwhat Gemini Robotics is."
  },
  {
    "start": 494.327,
    "end": 495.995,
    "text": "And then we continuously improve it."
  },
  {
    "start": 495.995,
    "end": 498.08,
    "text": "And Gemini Robotics 2 is a pretty big"
  },
  {
    "start": 498.08,
    "end": 500.917,
    "text": "step function with\nrespect to our previous one."
  },
  {
    "start": 500.917,
    "end": 501.709,
    "text": "- I love that."
  },
  {
    "start": 501.709,
    "end": 502.502,
    "text": "- I think, yeah, our team"
  },
  {
    "start": 502.502,
    "end": 503.628,
    "text": "aggressively thinks about this problem"
  },
  {
    "start": 503.628,
    "end": 506.047,
    "text": "from a frontier model, lab perspective."
  },
  {
    "start": 506.047,
    "end": 507.715,
    "text": "So, really leveraging,"
  },
  {
    "start": 507.715,
    "end": 508.424,
    "text": "you don't have to"
  },
  {
    "start": 508.424,
    "end": 509.675,
    "text": "solve robotics from scratch."
  },
  {
    "start": 509.675,
    "end": 511.177,
    "text": "There's all this world understanding"
  },
  {
    "start": 511.177,
    "end": 512.47,
    "text": "in these big models."
  },
  {
    "start": 512.47,
    "end": 513.971,
    "text": "So, the idea is how do you"
  },
  {
    "start": 513.971,
    "end": 515.681,
    "text": "latch on to that understanding"
  },
  {
    "start": 515.681,
    "end": 518.017,
    "text": "and then use the\nrobot for useful things?"
  },
  {
    "start": 518.017,
    "end": 521.812,
    "text": "So, you've had this history of using\ndata as a scaling paradigm, but now,"
  },
  {
    "start": 521.812,
    "end": 525.316,
    "text": "it's more like these\nfrontier intelligent things,"
  },
  {
    "start": 525.316,
    "end": 526.4,
    "text": "how do you hook up this"
  },
  {
    "start": 526.4,
    "end": 528.402,
    "text": "physical thing to those models and then"
  },
  {
    "start": 528.402,
    "end": 530.988,
    "text": "kind of bootstrap robotics from those?"
  },
  {
    "start": 530.988,
    "end": 533.324,
    "text": "- And what ends up being\nthe limitation in practice?"
  },
  {
    "start": 533.324,
    "end": 536.536,
    "text": "Because I'm thinking, for example,\nI've seen the demos of the robots"
  },
  {
    "start": 536.536,
    "end": 539.83,
    "text": "dancing and all this crazy stuff,\nand doing backflips, which I cannot do."
  },
  {
    "start": 539.83,
    "end": 540.957,
    "text": "But then you also see"
  },
  {
    "start": 540.957,
    "end": 545.67,
    "text": "for AI agents,\nthe meme canonical demo is like booking travel."
  },
  {
    "start": 545.67,
    "end": 549.382,
    "text": "And I feel like for robots,\nit's folding laundry."
  },
  {
    "start": 549.382,
    "end": 552.677,
    "text": "But yet, it feels like also maybe the\nrobots cannot actually fold laundry yet."
  },
  {
    "start": 552.677,
    "end": 556.43,
    "text": "Maybe that's not true.\nMaybe Gemini Robotics 2 closes that gap."
  },
  {
    "start": 556.43,
    "end": 559.016,
    "text": "But I'm curious,\nthere's so much understanding"
  },
  {
    "start": 559.016,
    "end": 562.979,
    "text": "and intelligence baked into the model."
  },
  {
    "start": 562.979,
    "end": 565.147,
    "text": "What's the place to actually hill climb"
  },
  {
    "start": 565.147,
    "end": 568.15,
    "text": "to get to the place\nwhere you start to see"
  },
  {
    "start": 568.15,
    "end": 571.32,
    "text": "robots folding laundry\nsuccessfully in most cases?"
  },
  {
    "start": 571.32,
    "end": 573.03,
    "text": "Or actually are we already there?"
  },
  {
    "start": 573.03,
    "end": 576.534,
    "text": "And maybe I'm not fully calibrated\non model progress in this regard."
  },
  {
    "start": 577.034,
    "end": 578.244,
    "text": "- So, there is a lot of"
  },
  {
    "start": 578.244,
    "end": 579.745,
    "text": "progress in the last ten years."
  },
  {
    "start": 579.745,
    "end": 580.997,
    "text": "So, one thing that you mentioned"
  },
  {
    "start": 580.997,
    "end": 583.249,
    "text": "is locomotion or whole body control"
  },
  {
    "start": 583.249,
    "end": 584.917,
    "text": "has reached a new height"
  },
  {
    "start": 584.917,
    "end": 586.294,
    "text": "where you see the humanoid robots"
  },
  {
    "start": 586.294,
    "end": 587.795,
    "text": "that can backflip and"
  },
  {
    "start": 587.795,
    "end": 589.255,
    "text": "doing very agile motions."
  },
  {
    "start": 589.255,
    "end": 593.05,
    "text": "Also, our team pioneered a\nresearch called reinforced learning"
  },
  {
    "start": 593.05,
    "end": 596.178,
    "text": "and sim-to-real transfer that\nmakes all those things happen."
  },
  {
    "start": 596.178,
    "end": 599.39,
    "text": "So, it seems that the\nlocomotion is nearly a solved problem."
  },
  {
    "start": 599.39,
    "end": 603.144,
    "text": "What's remaining is actually a very hard problem,\nis dexterous manipulation,"
  },
  {
    "start": 603.144,
    "end": 606.856,
    "text": "which is that how can you use\na hand or grippers to interact"
  },
  {
    "start": 606.856,
    "end": 610.943,
    "text": "with all these objects in order to\naccomplish the task in your daily life?"
  },
  {
    "start": 610.943,
    "end": 611.235,
    "text": "- Yeah."
  },
  {
    "start": 611.235,
    "end": 615.281,
    "text": "- The reason that's very complicated\nis because it's very contact-rich."
  },
  {
    "start": 615.281,
    "end": 619.16,
    "text": "So you need to think\nabout a lot of contact points"
  },
  {
    "start": 619.16,
    "end": 622.371,
    "text": "on the object in order to\nmove them in a desirable way."
  },
  {
    "start": 622.371,
    "end": 627.168,
    "text": "And to control your hand,\nit has over twenty degrees of freedom."
  },
  {
    "start": 627.168,
    "end": 629.545,
    "text": "So, you need to coordinate\nall these joints and muscles"
  },
  {
    "start": 629.545,
    "end": 631.922,
    "text": "in order to do those tasks."
  },
  {
    "start": 631.922,
    "end": 635.384,
    "text": "And all these things are way, way\nharder than locomotion problems."
  },
  {
    "start": 635.384,
    "end": 638.429,
    "text": "You only need to control yourself,\non usually a flat ground,"
  },
  {
    "start": 638.429,
    "end": 640.431,
    "text": "or slightly perturbed ground."
  },
  {
    "start": 640.431,
    "end": 643.267,
    "text": "So, dexterous manipulation\nwith all these different objects"
  },
  {
    "start": 643.267,
    "end": 647.688,
    "text": "in real world is really an\nunsolved problem for robotics for now."
  },
  {
    "start": 647.688,
    "end": 648.773,
    "text": "- Interesting."
  },
  {
    "start": 648.773,
    "end": 651.15,
    "text": "Where is the quality\ngain come from these days?"
  },
  {
    "start": 651.15,
    "end": 654.195,
    "text": "Is it we just get more data or it's"
  },
  {
    "start": 654.195,
    "end": 656.906,
    "text": "new techniques or just scaling\nup general purpose models?"
  },
  {
    "start": 656.906,
    "end": 658.783,
    "text": "How do we actually,"
  },
  {
    "start": 658.783,
    "end": 661.035,
    "text": "it's an unsolved problem,\nbut where do we make"
  },
  {
    "start": 661.035,
    "end": 663.788,
    "text": "progress on actually solving it?"
  },
  {
    "start": 663.788,
    "end": 665.456,
    "text": "- I think data is a big part of it."
  },
  {
    "start": 665.456,
    "end": 669.627,
    "text": "As Jie mentioned, we're missing this\ninternet of physical interaction data."
  },
  {
    "start": 669.627,
    "end": 671.962,
    "text": "As I open this cup,\nthere's a sequence of,"
  },
  {
    "start": 671.962,
    "end": 675.424,
    "text": "I make a move and the\nenvironment moves in response to it."
  },
  {
    "start": 675.424,
    "end": 678.052,
    "text": "So, this kind of sequence of\ninteraction with the physical world,"
  },
  {
    "start": 678.052,
    "end": 679.845,
    "text": "there's no internet of this."
  },
  {
    "start": 679.845,
    "end": 684.517,
    "text": "So, I think data is a key\npart to this physical AGI component"
  },
  {
    "start": 684.517,
    "end": 685.768,
    "text": "that we have not unlocked yet."
  },
  {
    "start": 685.768,
    "end": 688.896,
    "text": "And that is an open question,\nhow do you collect it?"
  },
  {
    "start": 688.896,
    "end": 690.606,
    "text": "Data quality really matters here."
  },
  {
    "start": 690.606,
    "end": 693.317,
    "text": "So, I think, yeah, just\nhow do we collect the scaled"
  },
  {
    "start": 693.317,
    "end": 695.945,
    "text": "digital version of physical\ninteractions is an open thing."
  },
  {
    "start": 695.945,
    "end": 698.072,
    "text": "And we're basically hill climbing that"
  },
  {
    "start": 698.072,
    "end": 701.075,
    "text": "as one of the big levers\non unlocking physical AGI."
  },
  {
    "start": 701.075,
    "end": 703.828,
    "text": "- Do we not have the ability to,\nand I'm guessing, actually,"
  },
  {
    "start": 703.828,
    "end": 707.289,
    "text": "one of the threads that I'm\nalways talking to folks about is"
  },
  {
    "start": 707.289,
    "end": 710.918,
    "text": "back to this building on the world model,\nthis interop between main Gemini"
  },
  {
    "start": 710.918,
    "end": 713.504,
    "text": "and some of these domain-specific cases."
  },
  {
    "start": 713.504,
    "end": 719.51,
    "text": "Can you not take a video of someone\nunscrewing the water bottle and intuit"
  },
  {
    "start": 719.51,
    "end": 722.972,
    "text": "some of this data and get\nsome of it out of those types of,"
  },
  {
    "start": 722.972,
    "end": 726.225,
    "text": "and it feels like there's a\nrichness in that type of data."
  },
  {
    "start": 726.225,
    "end": 728.769,
    "text": "Maybe there's not\nrichness in the labeled format"
  },
  {
    "start": 728.769,
    "end": 731.147,
    "text": "that we actually need to make progress,"
  },
  {
    "start": 731.147,
    "end": 734.233,
    "text": "but have we gotten closer to\nbeing able to actually leverage"
  },
  {
    "start": 734.233,
    "end": 738.195,
    "text": "some of the existing\nnon-robotics data to do these tasks?"
  },
  {
    "start": 738.195,
    "end": 740.781,
    "text": "- I think if you look\nat the data progression"
  },
  {
    "start": 740.781,
    "end": 741.991,
    "text": "in the last couple of years,"
  },
  {
    "start": 741.991,
    "end": 744.702,
    "text": "so, people usually talk\nabout this data pyramid."
  },
  {
    "start": 744.702,
    "end": 747.705,
    "text": "On the top of that\nis teleoperation data."
  },
  {
    "start": 747.705,
    "end": 751.459,
    "text": "These are the data where\nyou move some controllers,"
  },
  {
    "start": 751.459,
    "end": 753.419,
    "text": "and the robot is\ngoing to move accordingly,"
  },
  {
    "start": 753.419,
    "end": 755.463,
    "text": "so, this is called teleoperation."
  },
  {
    "start": 755.463,
    "end": 758.924,
    "text": "These data are very useful\nbecause you get all the signals,"
  },
  {
    "start": 758.924,
    "end": 761.552,
    "text": "basically how you should move\nthe robot in different scenarios"
  },
  {
    "start": 761.552,
    "end": 762.762,
    "text": "to train the robot."
  },
  {
    "start": 762.762,
    "end": 766.557,
    "text": "But, these data are not very scalable\nbecause teleoperation is very costly."
  },
  {
    "start": 766.557,
    "end": 769.351,
    "text": "You need a human there,\nyou need a robot in the loop, and so on."
  },
  {
    "start": 769.351,
    "end": 772.438,
    "text": "And people say,\n\"Maybe we need something more scalable.\""
  },
  {
    "start": 772.438,
    "end": 774.857,
    "text": "And people think about wearable device."
  },
  {
    "start": 774.857,
    "end": 778.194,
    "text": "There is something called the UMI,\nwhich is in the academic world,"
  },
  {
    "start": 778.194,
    "end": 783.365,
    "text": "where people build these\nwearable grippers so that human"
  },
  {
    "start": 783.365,
    "end": 785.618,
    "text": "can collect the data\nwithout the robot in the loop."
  },
  {
    "start": 785.618,
    "end": 787.328,
    "text": "- Pretend you're a robot in your house,"
  },
  {
    "start": 787.328,
    "end": 788.579,
    "text": "and do stuff all day."
  },
  {
    "start": 788.579,
    "end": 791.04,
    "text": "- I've seen some of these videos,\nit's interesting."
  },
  {
    "start": 791.04,
    "end": 793.918,
    "text": "- Then, those data\nbecomes really scalable."
  },
  {
    "start": 793.918,
    "end": 796.045,
    "text": "But the problem is human\nand robots are different."
  },
  {
    "start": 796.045,
    "end": 799.465,
    "text": "There is this embodiment\ngap you need to cross, right?"
  },
  {
    "start": 799.465,
    "end": 800.758,
    "text": "And down beneath it,"
  },
  {
    "start": 800.758,
    "end": 804.303,
    "text": "maybe the widest base is\nreally egocentric human data."
  },
  {
    "start": 804.303,
    "end": 806.931,
    "text": "Basically, human—\nyou take a video of human doing things,"
  },
  {
    "start": 806.931,
    "end": 808.891,
    "text": "then hopefully we can learn from those."
  },
  {
    "start": 808.891,
    "end": 812.102,
    "text": "But again,\nyou don't know how much actuation"
  },
  {
    "start": 812.102,
    "end": 816.482,
    "text": "or how much muscle force you\ndo with each movement of humans."
  },
  {
    "start": 816.482,
    "end": 820.152,
    "text": "So you are missing a very\nimportant label of the actions."
  },
  {
    "start": 820.152,
    "end": 823.155,
    "text": "At the same time, as I mentioned,\nthat human robots are different,"
  },
  {
    "start": 823.155,
    "end": 826.033,
    "text": "so it's the hardest data to leverage."
  },
  {
    "start": 826.033,
    "end": 828.41,
    "text": "Of course, we are\nmaking progress on leveraging"
  },
  {
    "start": 828.41,
    "end": 832.373,
    "text": "all the entire pyramid of data,\nbut we are not quite there yet."
  },
  {
    "start": 832.373,
    "end": 835.417,
    "text": "- Yeah, I'm thinking about in\nthe context of mainline Gemini,"
  },
  {
    "start": 835.417,
    "end": 838.003,
    "text": "for some of these use cases where\nlike the models don't really work,"
  },
  {
    "start": 838.003,
    "end": 841.549,
    "text": "you start to see like\n1,000 high-quality trajectories,"
  },
  {
    "start": 841.549,
    "end": 844.51,
    "text": "it makes a massive difference\nin overall quality of the model."
  },
  {
    "start": 844.51,
    "end": 848.18,
    "text": "Do you see that type of, again,\nwithout getting into the specifics,"
  },
  {
    "start": 848.18,
    "end": 851.267,
    "text": "is it low-hanging fruit,\nfree hill climbing everywhere?"
  },
  {
    "start": 851.267,
    "end": 854.019,
    "text": "Or is it really you\nactually need the reason tele--"
  },
  {
    "start": 854.019,
    "end": 855.771,
    "text": "because I'm thinking, it's Google."
  },
  {
    "start": 855.771,
    "end": 860.276,
    "text": "We could we could go get\n1,000 tele-operated,"
  },
  {
    "start": 860.276,
    "end": 862.653,
    "text": "examples of people going in,\ncollecting that data."
  },
  {
    "start": 862.653,
    "end": 864.28,
    "text": "But you're saying we need it would be"
  },
  {
    "start": 864.28,
    "end": 867.867,
    "text": "the scales of millions of\nmillions in order to get any..."
  },
  {
    "start": 867.867,
    "end": 868.826,
    "text": "Interesting."
  },
  {
    "start": 868.826,
    "end": 869.41,
    "text": "- This was Gemini,"
  },
  {
    "start": 869.41,
    "end": 870.578,
    "text": "which is pre-trained\non the internet, right?"
  },
  {
    "start": 870.578,
    "end": 874.999,
    "text": "So there's a lot of nice biases\nthere for language and vision stuff."
  },
  {
    "start": 874.999,
    "end": 875.416,
    "text": "- Yeah."
  },
  {
    "start": 875.416,
    "end": 878.335,
    "text": "- But whenever we add this\nphysical thing into the model,"
  },
  {
    "start": 878.335,
    "end": 880.504,
    "text": "it doesn’t play well\nwith the pre-trained stuff."
  },
  {
    "start": 880.504,
    "end": 882.172,
    "text": "So in fact,\nthat's one of the reasons why we have"
  },
  {
    "start": 882.172,
    "end": 885.342,
    "text": "our Gemini Robotics model is\nbecause it is hard to upstream something"
  },
  {
    "start": 885.342,
    "end": 888.053,
    "text": "without killing all the other\ngenerational properties of the model."
  },
  {
    "start": 888.053,
    "end": 888.679,
    "text": "So,"
  },
  {
    "start": 888.679,
    "end": 891.39,
    "text": "this physical thing,\nour datasets are so tiny compared to the other"
  },
  {
    "start": 891.39,
    "end": 894.435,
    "text": "digital token sets,\nthat they don't play well yet."
  },
  {
    "start": 894.435,
    "end": 896.604,
    "text": "So, either we scale these up\nand they start playing well,"
  },
  {
    "start": 896.604,
    "end": 898.022,
    "text": "or there's some other\nway we can connect them."
  },
  {
    "start": 898.022,
    "end": 899.315,
    "text": "But it's still an unsolved problem."
  },
  {
    "start": 899.315,
    "end": 901.859,
    "text": "That's why you don't\nsee these frontier models"
  },
  {
    "start": 901.859,
    "end": 903.235,
    "text": "directly controlling robots."
  },
  {
    "start": 903.235,
    "end": 906.447,
    "text": "- Yeah. You inherit things like\nnatural language understanding,"
  },
  {
    "start": 906.447,
    "end": 907.907,
    "text": "visual understanding."
  },
  {
    "start": 907.907,
    "end": 912.328,
    "text": "I don't have to pick up a bottle that\nis black and white and different shapes,"
  },
  {
    "start": 912.328,
    "end": 915.039,
    "text": "like all of that generalization,\nwhich we now take for granted."
  },
  {
    "start": 915.039,
    "end": 916.957,
    "text": "- That's true.\nI did not even think about that."
  },
  {
    "start": 916.957,
    "end": 919.877,
    "text": "- But before you had to\ncollect every single object."
  },
  {
    "start": 919.877,
    "end": 921.253,
    "text": "So, we do get some generalization,"
  },
  {
    "start": 921.253,
    "end": 923.422,
    "text": "but I think the big part that is missing"
  },
  {
    "start": 923.422,
    "end": 925.007,
    "text": "is really understanding motion."
  },
  {
    "start": 925.007,
    "end": 928.969,
    "text": "And motion is not something that\nyou inherit from Gemini today, right?"
  },
  {
    "start": 928.969,
    "end": 931.347,
    "text": "Motion is something that\nwe have to teach Gemini."
  },
  {
    "start": 931.347,
    "end": 932.64,
    "text": "- Or understand force."
  },
  {
    "start": 932.64,
    "end": 933.849,
    "text": "- Or force.\n- Interactions."
  },
  {
    "start": 933.849,
    "end": 936.06,
    "text": "These are the things that I\nthink Gemini is not trained on."
  },
  {
    "start": 936.06,
    "end": 940.022,
    "text": "- Yeah. I'm also super curious,\nthere's obviously such a distribution"
  },
  {
    "start": 940.022,
    "end": 941.732,
    "text": "of the actual robotic use cases."
  },
  {
    "start": 941.732,
    "end": 944.485,
    "text": "And I'm curious, for us,\nand for Gemini Robotics,"
  },
  {
    "start": 944.485,
    "end": 945.819,
    "text": "has there been a focus?"
  },
  {
    "start": 945.819,
    "end": 951.408,
    "text": "Is it like we want to enable home\nrobotic use cases and we see traction?"
  },
  {
    "start": 951.408,
    "end": 954.161,
    "text": "Obviously, there's a huge amount of\nindustrial automation stuff happening."
  },
  {
    "start": 954.161,
    "end": 957.581,
    "text": "Is there-- I know, Carolina,\nyou said we want general purpose robots."
  },
  {
    "start": 957.581,
    "end": 959.625,
    "text": "And so theoretically you\ncould do all of those things,"
  },
  {
    "start": 959.625,
    "end": 960.417,
    "text": "but I'm curious, actually,"
  },
  {
    "start": 960.417,
    "end": 965.714,
    "text": "if there's been-- is it jagged\nas far as progress or capability,"
  },
  {
    "start": 965.714,
    "end": 970.219,
    "text": "or things that are actually\nworking more today versus not?"
  },
  {
    "start": 970.219,
    "end": 972.68,
    "text": "- Yeah, there is, I think,\ntwo answers to that question."
  },
  {
    "start": 972.68,
    "end": 976.1,
    "text": "I would say in the capability front,\nwe certainly believe that you want to"
  },
  {
    "start": 976.1,
    "end": 980.145,
    "text": "be able to solve a broad range of tasks,\nand actually by going narrow,"
  },
  {
    "start": 980.145,
    "end": 982.898,
    "text": "you're going to build a policy\nthat is going to be a lot more brittle."
  },
  {
    "start": 982.898,
    "end": 985.442,
    "text": "It might work better in that environment,\nbut the minute you change"
  },
  {
    "start": 985.442,
    "end": 988.07,
    "text": "anything about the environment,\nthings are going to start to go wrong."
  },
  {
    "start": 988.07,
    "end": 991.407,
    "text": "So, our approach is\ncertainly don't compromise on,"
  },
  {
    "start": 991.407,
    "end": 993.492,
    "text": "you're trying to solve\ngeneral-purpose tasks,"
  },
  {
    "start": 993.492,
    "end": 995.285,
    "text": "and understand general-purpose motion,"
  },
  {
    "start": 995.285,
    "end": 997.621,
    "text": "and handle a broad\nrange of different objects."
  },
  {
    "start": 997.621,
    "end": 1000.874,
    "text": "I would say in terms of deployment,\nI think of it a bit differently, right?"
  },
  {
    "start": 1000.874,
    "end": 1004.253,
    "text": "I think it's much more likely that\nthese robots are going to be useful"
  },
  {
    "start": 1004.253,
    "end": 1007.631,
    "text": "in environments like industrial\nenvironments that are semi-structured,"
  },
  {
    "start": 1007.631,
    "end": 1013.012,
    "text": "that have safety pretty much under control,\nand that you can get"
  },
  {
    "start": 1013.012,
    "end": 1014.304,
    "text": "a lot of real-world experience"
  },
  {
    "start": 1014.304,
    "end": 1016.265,
    "text": "of what it means to launch these models."
  },
  {
    "start": 1016.265,
    "end": 1018.517,
    "text": "And then from there,\nI imagine that we would go to things"
  },
  {
    "start": 1018.517,
    "end": 1021.395,
    "text": "like retail and other\nenvironments that are also"
  },
  {
    "start": 1021.395,
    "end": 1025.649,
    "text": "starting to get into human-centric spaces,\nbut they're less vulnerable"
  },
  {
    "start": 1025.649,
    "end": 1028.61,
    "text": "than in the middle of your house,\nwith your kids and your pets."
  },
  {
    "start": 1028.61,
    "end": 1033.574,
    "text": "- You're raining on my Q4\n2026 home robot orders right now."
  },
  {
    "start": 1033.574,
    "end": 1034.116,
    "text": "I'm waiting."
  },
  {
    "start": 1034.116,
    "end": 1037.411,
    "text": "I've written off all of\nmy chores for 2026,"
  },
  {
    "start": 1037.411,
    "end": 1040.414,
    "text": "as soon as I start getting\nsome of these robots delivered."
  },
  {
    "start": 1040.414,
    "end": 1041.373,
    "text": "So it's not—"
  },
  {
    "start": 1041.373,
    "end": 1042.958,
    "text": "- Yeah, I mean,\nthere's plenty of people out there"
  },
  {
    "start": 1042.958,
    "end": 1044.543,
    "text": "that think that they're\ngoing to go home first."
  },
  {
    "start": 1044.543,
    "end": 1046.67,
    "text": "And I think there's,\nthe appealing thing there"
  },
  {
    "start": 1046.67,
    "end": 1050.674,
    "text": "is that home first requires that diversity,\nthat generalization."
  },
  {
    "start": 1050.674,
    "end": 1054.595,
    "text": "So it forces that problem,\nit makes it very front and center."
  },
  {
    "start": 1054.595,
    "end": 1056.18,
    "text": "- And you have to check,\nI think we always found"
  },
  {
    "start": 1056.18,
    "end": 1059.808,
    "text": "it's the opposite where,\nif you do train on this one narrow thing,"
  },
  {
    "start": 1059.808,
    "end": 1061.435,
    "text": "as yeah, it becomes worse."
  },
  {
    "start": 1061.435,
    "end": 1064.563,
    "text": "So having it collect on\nmany different diverse cases,"
  },
  {
    "start": 1064.563,
    "end": 1066.148,
    "text": "just helps the\ngeneral intelligence of it."
  },
  {
    "start": 1066.148,
    "end": 1069.86,
    "text": "So yeah, from a strategy perspective\nand a learning perspective,"
  },
  {
    "start": 1069.86,
    "end": 1072.196,
    "text": "it makes sense to go super broad first."
  },
  {
    "start": 1072.196,
    "end": 1072.696,
    "text": "- Yeah."
  },
  {
    "start": 1072.696,
    "end": 1074.656,
    "text": "- I mean, then there could be\nthe super fans, like yourself,"
  },
  {
    "start": 1074.656,
    "end": 1078.952,
    "text": "that decide to have the robot at home,\neven if it's a little early."
  },
  {
    "start": 1078.952,
    "end": 1081.08,
    "text": "- I hope it doesn't break my stuff."
  },
  {
    "start": 1081.08,
    "end": 1084.5,
    "text": "Yeah, I think it is interesting to\nsee like what people's— because I assume"
  },
  {
    "start": 1084.5,
    "end": 1086.877,
    "text": "these robots will be\ndelivered to people and actually,"
  },
  {
    "start": 1086.877,
    "end": 1088.629,
    "text": "I think it'll be super\ninteresting for all of us to"
  },
  {
    "start": 1088.629,
    "end": 1090.547,
    "text": "just see what's people's reaction."
  },
  {
    "start": 1090.547,
    "end": 1094.176,
    "text": "If it works 90% of the\ntime and 10% of the time it's,"
  },
  {
    "start": 1094.176,
    "end": 1096.47,
    "text": "you know, cracking a wine glass\nor something like that."
  },
  {
    "start": 1096.47,
    "end": 1098.305,
    "text": "Are you happy with that?"
  },
  {
    "start": 1098.305,
    "end": 1101.35,
    "text": "If it's a $3 wine glass,\nmaybe, I have no idea."
  },
  {
    "start": 1101.35,
    "end": 1105.771,
    "text": "Part of this challenge of scaling up data,\nI'm curious why,"
  },
  {
    "start": 1105.771,
    "end": 1109.274,
    "text": "actually, back to this home example,\nand maybe there's a bunch of industrial examples,"
  },
  {
    "start": 1109.274,
    "end": 1112.361,
    "text": "the tension and you see those--\nopposite extreme of this"
  },
  {
    "start": 1112.361,
    "end": 1114.279,
    "text": "in the context of mainline Gemini where,"
  },
  {
    "start": 1114.279,
    "end": 1116.615,
    "text": "you know, there's billions of people using,\net cetera, et cetera."
  },
  {
    "start": 1116.615,
    "end": 1119.91,
    "text": "We sort of have a flywheel of\ngetting signal from the real world."
  },
  {
    "start": 1119.91,
    "end": 1127.459,
    "text": "Why don't we see more real live\ndeployments of robots today in,"
  },
  {
    "start": 1127.459,
    "end": 1130.254,
    "text": "actually, to help us scale\ngetting a lot of this data"
  },
  {
    "start": 1130.254,
    "end": 1133.382,
    "text": "and some of these ways,\nis just like it's not scalable?"
  },
  {
    "start": 1133.382,
    "end": 1135.217,
    "text": "And I'm thinking back to,\nand maybe this is not right,"
  },
  {
    "start": 1135.217,
    "end": 1138.47,
    "text": "But you see, obviously,\nWaymo was doing this for a long time,"
  },
  {
    "start": 1138.47,
    "end": 1140.848,
    "text": "and had cars driving around,\nnot being used."
  },
  {
    "start": 1140.848,
    "end": 1143.934,
    "text": "There's actually a ton of other\nself-driving car startups still doing"
  },
  {
    "start": 1143.934,
    "end": 1147.729,
    "text": "similar things today and in operation,\ntheoretically collecting data."
  },
  {
    "start": 1147.729,
    "end": 1148.981,
    "text": "I don't know what\nthey're actually doing."
  },
  {
    "start": 1148.981,
    "end": 1151.984,
    "text": "And maybe I'm wrong about this,\nbut it feels like that isn’t the case"
  },
  {
    "start": 1151.984,
    "end": 1154.736,
    "text": "of what is happening in robotics today,"
  },
  {
    "start": 1154.736,
    "end": 1157.447,
    "text": "at least visibly to\nan external observer."
  },
  {
    "start": 1157.447,
    "end": 1158.49,
    "text": "And I'm curious why."
  },
  {
    "start": 1158.49,
    "end": 1159.616,
    "text": "- Do you want to answer that?"
  },
  {
    "start": 1159.616,
    "end": 1162.035,
    "text": "- We are going to see over the\nnext two years if that changes a lot."
  },
  {
    "start": 1162.035,
    "end": 1164.329,
    "text": "But I think right now,\nwhen we talk to partners"
  },
  {
    "start": 1164.329,
    "end": 1165.581,
    "text": "about their own experiences,"
  },
  {
    "start": 1165.581,
    "end": 1168.125,
    "text": "we hear two things consistently."
  },
  {
    "start": 1168.125,
    "end": 1169.877,
    "text": "One is, while the policies are showing"
  },
  {
    "start": 1169.877,
    "end": 1172.838,
    "text": "incredibly cool generalizing capabilities,"
  },
  {
    "start": 1172.838,
    "end": 1174.089,
    "text": "it's actually difficult to get"
  },
  {
    "start": 1174.089,
    "end": 1177.05,
    "text": "them to be narrowly successful,"
  },
  {
    "start": 1177.05,
    "end": 1177.885,
    "text": "but broad enough"
  },
  {
    "start": 1177.885,
    "end": 1178.552,
    "text": "that they can handle"
  },
  {
    "start": 1178.552,
    "end": 1180.721,
    "text": "all the little things\nthat vary and break"
  },
  {
    "start": 1180.721,
    "end": 1183.515,
    "text": "over the course of an entire day.\nAnd I think you look back to autonomy."
  },
  {
    "start": 1183.515,
    "end": 1185.225,
    "text": "This is a really serious\nchallenge there as well,"
  },
  {
    "start": 1185.225,
    "end": 1187.311,
    "text": "where you can give a\nreally impressive demo,"
  },
  {
    "start": 1187.311,
    "end": 1189.98,
    "text": "but you're not ready to remove\nthe safety driver for a long time."
  },
  {
    "start": 1189.98,
    "end": 1190.48,
    "text": "- Yeah."
  },
  {
    "start": 1190.48,
    "end": 1193.442,
    "text": "- And one of the distinctions\nbetween autonomy and robotics"
  },
  {
    "start": 1193.442,
    "end": 1195.152,
    "text": "is this teleoperation option."
  },
  {
    "start": 1195.152,
    "end": 1197.487,
    "text": "So, you can just like,\n\"I'll just put a safety driver in.\""
  },
  {
    "start": 1197.487,
    "end": 1200.324,
    "text": "And if the car gets stuck,\nthey'll just take over and drive."
  },
  {
    "start": 1200.324,
    "end": 1203.16,
    "text": "For a lot of the robots,\nyou know, we saw this with Aloha."
  },
  {
    "start": 1203.16,
    "end": 1203.785,
    "text": "We could do that."
  },
  {
    "start": 1203.785,
    "end": 1205.078,
    "text": "We could actually get\nsomebody right there."
  },
  {
    "start": 1205.078,
    "end": 1207.873,
    "text": "And if the robot got stuck,\nthey could take over and fix it."
  },
  {
    "start": 1207.873,
    "end": 1210.751,
    "text": "But as the robots get bigger,\nget more capable,"
  },
  {
    "start": 1210.751,
    "end": 1213.962,
    "text": "gain more degrees of freedom,\nwhich we should actually want in a deployment,"
  },
  {
    "start": 1213.962,
    "end": 1216.673,
    "text": "it's harder and harder for\nsomebody to jump in and help."
  },
  {
    "start": 1216.673,
    "end": 1218.3,
    "text": "And so, I think we're\nseeing this challenge right now,"
  },
  {
    "start": 1218.3,
    "end": 1221.595,
    "text": "where we're struggling to\nget that bootstrap system"
  },
  {
    "start": 1221.595,
    "end": 1224.389,
    "text": "where it's good enough\nthat you're ready to deploy it."
  },
  {
    "start": 1224.389,
    "end": 1227.142,
    "text": "And then also we're trying to\nfigure out what is this analogy where"
  },
  {
    "start": 1227.142,
    "end": 1228.644,
    "text": "if it does get stuck,"
  },
  {
    "start": 1228.644,
    "end": 1231.772,
    "text": "how do I both fix it quickly\nand learn from that moment?"
  },
  {
    "start": 1231.772,
    "end": 1235.567,
    "text": "And so there's little patterns forming,\nbut right now I think everyone"
  },
  {
    "start": 1235.567,
    "end": 1238.862,
    "text": "is really focused on how do I get\nthis kind of baseline performance?"
  },
  {
    "start": 1238.862,
    "end": 1240.405,
    "text": "And so I think there's a\nbig rush in the industry"
  },
  {
    "start": 1240.405,
    "end": 1242.282,
    "text": "and I get that baseline capability."
  },
  {
    "start": 1242.282,
    "end": 1242.991,
    "text": "- Yeah."
  },
  {
    "start": 1242.991,
    "end": 1245.285,
    "text": "- And this is part of what we're\ndoing with our partners actually,"
  },
  {
    "start": 1245.285,
    "end": 1247.871,
    "text": "so, as you notice, there's many\ndifferent robot types here,"
  },
  {
    "start": 1247.871,
    "end": 1249.623,
    "text": "and none of them were\nbuilt by us actually."
  },
  {
    "start": 1249.623,
    "end": 1249.706,
    "text": "- Yeah."
  },
  {
    "start": 1249.706,
    "end": 1252.876,
    "text": "- So the way we work is that we have\nthese deep partners that we work with"
  },
  {
    "start": 1252.876,
    "end": 1257.297,
    "text": "in order to accelerate both the\nAI capabilities and the hardware."
  },
  {
    "start": 1257.297,
    "end": 1259.883,
    "text": "And so we're deeply\nconnecting with each other"
  },
  {
    "start": 1259.883,
    "end": 1262.636,
    "text": "to see what's missing\nto get to that deployment."
  },
  {
    "start": 1262.636,
    "end": 1265.555,
    "text": "And our goals together is\nto accelerate that deployment"
  },
  {
    "start": 1265.555,
    "end": 1268.517,
    "text": "as fast as possible and to\nbring it to real-world applications."
  },
  {
    "start": 1268.517,
    "end": 1270.56,
    "text": "So that we can learn whether\nwhat we're learning is actually"
  },
  {
    "start": 1270.56,
    "end": 1274.356,
    "text": "useful and valuable and\nwhere to spend more time next."
  },
  {
    "start": 1274.356,
    "end": 1276.692,
    "text": "But that's absolutely the goal\nthat we have with our partners"
  },
  {
    "start": 1276.692,
    "end": 1279.861,
    "text": "is how to bring this to useful\napplications as soon as possible"
  },
  {
    "start": 1279.861,
    "end": 1283.323,
    "text": "and how to learn from it,\nhow to deploy them in a safe manner"
  },
  {
    "start": 1283.323,
    "end": 1287.077,
    "text": "so that we can get as much\ninformation as possible early on."
  },
  {
    "start": 1287.077,
    "end": 1289.83,
    "text": "- Yeah. Back to this two year time horizon,"
  },
  {
    "start": 1289.83,
    "end": 1291.665,
    "text": "potentially of where\nwe see things changing."
  },
  {
    "start": 1291.665,
    "end": 1296.003,
    "text": "I always make this comment to people that if you,\nten years ago,"
  },
  {
    "start": 1296.003,
    "end": 1299.923,
    "text": "were put in a very short-term\ntime machine and landed in 2026."
  },
  {
    "start": 1299.923,
    "end": 1303.302,
    "text": "If you looked outside,\nminus anywhere where there's Waymos deployed,"
  },
  {
    "start": 1303.302,
    "end": 1305.721,
    "text": "and you looked into a city,"
  },
  {
    "start": 1305.721,
    "end": 1307.472,
    "text": "essentially, physically,\neverything looks the same."
  },
  {
    "start": 1307.472,
    "end": 1308.515,
    "text": "like you wouldn't be able to--"
  },
  {
    "start": 1308.515,
    "end": 1311.226,
    "text": "maybe you'd spot a new phone that\nsomebody had or something like that,"
  },
  {
    "start": 1311.226,
    "end": 1312.978,
    "text": "but more or less,"
  },
  {
    "start": 1312.978,
    "end": 1315.314,
    "text": "the physical world\naround us looks the same."
  },
  {
    "start": 1315.314,
    "end": 1317.649,
    "text": "And they'd miss the fact\nthat we actually have these"
  },
  {
    "start": 1317.649,
    "end": 1321.361,
    "text": "extremely intelligent,\nwe've almost--not actually solved intelligence"
  },
  {
    "start": 1321.361,
    "end": 1324.156,
    "text": "but made a huge amount of\nprogress on solving intelligence."
  },
  {
    "start": 1324.156,
    "end": 1326.45,
    "text": "And so it's going to be\nvery interesting to see"
  },
  {
    "start": 1326.45,
    "end": 1329.703,
    "text": "the physical world around us\nstart to change as like you get these"
  },
  {
    "start": 1329.703,
    "end": 1334.249,
    "text": "new autonomous systems\ndoing things in the real world."
  },
  {
    "start": 1334.249,
    "end": 1336.918,
    "text": "I'm actually curious,\nassuming that we get some breakthroughs"
  },
  {
    "start": 1336.918,
    "end": 1339.296,
    "text": "the sort of dexterous\nmanipulation problem is solved."
  },
  {
    "start": 1339.296,
    "end": 1343.633,
    "text": "Is it basically a\nmanufacturing problem then,"
  },
  {
    "start": 1343.633,
    "end": 1345.51,
    "text": "at that point, to actually just scale--"
  },
  {
    "start": 1345.51,
    "end": 1351.308,
    "text": "or is there still fifty other things\nthat need to land if we were to somehow,"
  },
  {
    "start": 1351.308,
    "end": 1354.811,
    "text": "we could make dexterous manipulation\nwork really, really well."
  },
  {
    "start": 1354.811,
    "end": 1356.646,
    "text": "Everything out-- navigation is solved,"
  },
  {
    "start": 1356.646,
    "end": 1359.066,
    "text": "all the other bits\nof the story have been—"
  },
  {
    "start": 1359.066,
    "end": 1360.65,
    "text": "- Yeah, I mean, I think— first of all,"
  },
  {
    "start": 1360.65,
    "end": 1363.487,
    "text": "dexterous manipulation is\nthe hardest task problem."
  },
  {
    "start": 1363.487,
    "end": 1366.531,
    "text": "- So, we would all be very\nhappy when that is solved."
  },
  {
    "start": 1366.531,
    "end": 1368.283,
    "text": "But my guess is that if you want to"
  },
  {
    "start": 1368.283,
    "end": 1370.035,
    "text": "get to the point that\nI think we all dream up,"
  },
  {
    "start": 1370.035,
    "end": 1373.789,
    "text": "which is like, you walk around,\nand there's robots doing different things,"
  },
  {
    "start": 1373.789,
    "end": 1377.417,
    "text": "regardless of where you are,\nthey're helping society in a useful way."
  },
  {
    "start": 1377.417,
    "end": 1379.252,
    "text": "And not just in industrial settings,"
  },
  {
    "start": 1379.252,
    "end": 1381.88,
    "text": "but I think the second\nproblem we're going to hit"
  },
  {
    "start": 1381.88,
    "end": 1384.216,
    "text": "is the human-centric aspect of it."
  },
  {
    "start": 1384.216,
    "end": 1387.761,
    "text": "It's understanding humans,\nbeing able to be useful to humans,"
  },
  {
    "start": 1387.761,
    "end": 1390.138,
    "text": "and safe to humans in all that context."
  },
  {
    "start": 1390.138,
    "end": 1392.391,
    "text": "And that is something that\nwe're also making progress towards."
  },
  {
    "start": 1392.391,
    "end": 1395.06,
    "text": "I think for them,\nfor robots to be in everyday spaces,"
  },
  {
    "start": 1395.06,
    "end": 1398.73,
    "text": "there's all kinds of safety\naspects that need to also be solved,"
  },
  {
    "start": 1398.73,
    "end": 1402.109,
    "text": "all kinds of security,\nprivacy aspects that need to be solved"
  },
  {
    "start": 1402.109,
    "end": 1403.527,
    "text": "that are completely parallel."
  },
  {
    "start": 1403.527,
    "end": 1406.321,
    "text": "And very different to dexterity."
  },
  {
    "start": 1406.321,
    "end": 1411.159,
    "text": "But, I certainly think if we\ncrack the dexterity in a general way,"
  },
  {
    "start": 1411.159,
    "end": 1414.121,
    "text": "it would definitely\njust blow up the opportunity"
  },
  {
    "start": 1414.121,
    "end": 1415.163,
    "text": "of what's possible with robotics."
  },
  {
    "start": 1415.163,
    "end": 1418.291,
    "text": "- I think you will start seeing\nrobots around once we have that problem."
  },
  {
    "start": 1418.291,
    "end": 1421.128,
    "text": "We go to robotics conferences and\nwe get a little peek in the future."
  },
  {
    "start": 1421.128,
    "end": 1421.92,
    "text": "And it is crazy."
  },
  {
    "start": 1421.92,
    "end": 1424.089,
    "text": "You see these robots\nwalking around giving demos."
  },
  {
    "start": 1424.089,
    "end": 1427.509,
    "text": "So, it feels like Star Wars, the future."
  },
  {
    "start": 1427.509,
    "end": 1429.845,
    "text": "So, I feel like we're a\nfew years away from that."
  },
  {
    "start": 1429.845,
    "end": 1432.431,
    "text": "And this dexterous manipulation\nis one big chunk of that."
  },
  {
    "start": 1432.431,
    "end": 1435.142,
    "text": "So, maybe there's one or two of these,\nand then, yeah,"
  },
  {
    "start": 1435.142,
    "end": 1439.813,
    "text": "then the tail of the things that\nneeds to be solved, we'll get to, yeah,"
  },
  {
    "start": 1439.813,
    "end": 1442.983,
    "text": "we'll have robots around us,\nand maybe the cities will start looking different."
  },
  {
    "start": 1442.983,
    "end": 1446.903,
    "text": "- Yeah\n- I think robotics is incredibly hard."
  },
  {
    "start": 1446.903,
    "end": 1449.614,
    "text": "- Yeah. Make a t-shirt,\nrobotics is incredibly hard."
  },
  {
    "start": 1449.614,
    "end": 1452.617,
    "text": "- Yes. So, actually,\nI always get asked this question,"
  },
  {
    "start": 1452.617,
    "end": 1456.246,
    "text": "when do you feel robots is\ngoing to enter our daily life?"
  },
  {
    "start": 1456.246,
    "end": 1459.249,
    "text": "So, if you ask me like three years ago,\nI would say,"
  },
  {
    "start": 1459.249,
    "end": 1461.251,
    "text": "probably beyond my lifetime."
  },
  {
    "start": 1461.251,
    "end": 1461.918,
    "text": "- Interesting."
  },
  {
    "start": 1461.918,
    "end": 1466.465,
    "text": "- If you ask me two years ago,\nI said maybe ten years."
  },
  {
    "start": 1466.465,
    "end": 1468.3,
    "text": "- Is Waymo considered a robot or no?"
  },
  {
    "start": 1468.3,
    "end": 1470.635,
    "text": "- No. General-purpose robot."
  },
  {
    "start": 1470.635,
    "end": 1472.012,
    "text": "Enter our daily lives."
  },
  {
    "start": 1472.012,
    "end": 1473.763,
    "text": "So, if you ask me now,"
  },
  {
    "start": 1473.763,
    "end": 1476.099,
    "text": "I think that it's\nbetween five to ten years."
  },
  {
    "start": 1476.099,
    "end": 1481.688,
    "text": "So, you can see the speed of evolution\nof this technology is amazingly fast."
  },
  {
    "start": 1481.688,
    "end": 1483.815,
    "text": "But there are still a lot of\nthings that we need to solve."
  },
  {
    "start": 1484.191,
    "end": 1487.694,
    "text": "- I'm curious actually, though,\nlike five to ten years away framing."
  },
  {
    "start": 1487.694,
    "end": 1491.364,
    "text": "Also, if you stack that up on framing,"
  },
  {
    "start": 1491.364,
    "end": 1492.407,
    "text": "are we going to claim"
  },
  {
    "start": 1492.407,
    "end": 1495.869,
    "text": "general purpose intelligence if you\ndon't have this embodied characteristic?"
  },
  {
    "start": 1495.869,
    "end": 1499.873,
    "text": "Or are these two things\ncompletely separate?"
  },
  {
    "start": 1499.873,
    "end": 1501.458,
    "text": "- You have a biased group here."
  },
  {
    "start": 1501.458,
    "end": 1504.085,
    "text": "You cannot reach AGI\nuntil you solve physical AGI."
  },
  {
    "start": 1504.085,
    "end": 1507.631,
    "text": "So if I walked up to a robot and say,\ndo anything that I could do,"
  },
  {
    "start": 1507.631,
    "end": 1508.882,
    "text": "I would expect it to be able to do it."
  },
  {
    "start": 1508.882,
    "end": 1511.384,
    "text": "So I think that\ndefinition maybe sometimes"
  },
  {
    "start": 1511.384,
    "end": 1513.22,
    "text": "gets lost but we all\njust live it every day."
  },
  {
    "start": 1513.22,
    "end": 1515.931,
    "text": "So I think robotics,\nwe call this the Moravec's paradox,"
  },
  {
    "start": 1515.931,
    "end": 1519.935,
    "text": "where things that are really easy for\nhumans are very difficult for robots."
  },
  {
    "start": 1519.935,
    "end": 1523.271,
    "text": "Like, the AI is--\nit's passed the bar exam,"
  },
  {
    "start": 1523.271,
    "end": 1526.775,
    "text": "and code up all these operating systems,"
  },
  {
    "start": 1526.775,
    "end": 1529.736,
    "text": "but they cannot cook you eggs,\nor flip a burger."
  },
  {
    "start": 1529.736,
    "end": 1532.864,
    "text": "So, there's some paradox\nthere where physical AGI,"
  },
  {
    "start": 1532.864,
    "end": 1535.492,
    "text": "I think is a part of AGI,\nat least for me,"
  },
  {
    "start": 1535.492,
    "end": 1537.911,
    "text": "but it is in some ways\nmore fundamentally different"
  },
  {
    "start": 1537.911,
    "end": 1539.454,
    "text": "than the digital agents."
  },
  {
    "start": 1539.454,
    "end": 1543.25,
    "text": "So, it'll take, I think,\na bit more work to get that."
  },
  {
    "start": 1543.25,
    "end": 1546.753,
    "text": "I think that will land after\nthe digital AGI thing has happened."
  },
  {
    "start": 1546.753,
    "end": 1548.547,
    "text": "- Yeah.\n- So, I think there's,"
  },
  {
    "start": 1548.547,
    "end": 1552.592,
    "text": "that's why you add the two\nplus three to get to five."
  },
  {
    "start": 1552.592,
    "end": 1557.889,
    "text": "- Yeah, and I think it is very\npossible that getting to digital AGI,"
  },
  {
    "start": 1557.889,
    "end": 1560.016,
    "text": "and I'm sure it will, actually,\ndramatically accelerate"
  },
  {
    "start": 1560.016,
    "end": 1561.685,
    "text": "the speed to our physical AGI, right?"
  },
  {
    "start": 1561.685,
    "end": 1562.018,
    "text": "- Yeah."
  },
  {
    "start": 1562.018,
    "end": 1564.896,
    "text": "- Not only on the intelligent aspect,\nbut you could also use this"
  },
  {
    "start": 1564.896,
    "end": 1568.441,
    "text": "in order to build better robots,\non the hardware side."
  },
  {
    "start": 1568.441,
    "end": 1570.819,
    "text": "Because we haven't\ntalked much about sensors,"
  },
  {
    "start": 1570.819,
    "end": 1573.238,
    "text": "but everything that we're using today"
  },
  {
    "start": 1573.238,
    "end": 1577.659,
    "text": "is primarily ignoring all of the\nsensors that you have in your hand."
  },
  {
    "start": 1577.659,
    "end": 1583.79,
    "text": "So today, we're just using vision,\nand that's basically it,"
  },
  {
    "start": 1583.79,
    "end": 1585.875,
    "text": "and the position of the\nhand in order to determine"
  },
  {
    "start": 1585.875,
    "end": 1587.502,
    "text": "whether you have picked up this glass."
  },
  {
    "start": 1587.502,
    "end": 1591.339,
    "text": "But when I pick it up,\nI can feel it all over my hand."
  },
  {
    "start": 1591.339,
    "end": 1596.928,
    "text": "And that's something that we are not\neven scratching the surface on today."
  },
  {
    "start": 1596.928,
    "end": 1599.806,
    "text": "And we think that if you want to\nbe able to do everything a human can,"
  },
  {
    "start": 1599.806,
    "end": 1602.517,
    "text": "you definitely are going to need\nto have more sensing capabilities"
  },
  {
    "start": 1602.517,
    "end": 1604.144,
    "text": "than what we have today on robots."
  },
  {
    "start": 1604.144,
    "end": 1605.895,
    "text": "So, there's an aspect\nof also the hardware"
  },
  {
    "start": 1605.895,
    "end": 1607.48,
    "text": "catching up to getting to the level"
  },
  {
    "start": 1607.48,
    "end": 1613.486,
    "text": "that is capable of achieving\nhuman-level behaviors and manipulation."
  },
  {
    "start": 1613.486,
    "end": 1616.114,
    "text": "- Yeah, skin is an\nunsolved hardware problem."
  },
  {
    "start": 1616.114,
    "end": 1619.492,
    "text": "- Yeah, I can imagine that being true."
  },
  {
    "start": 1619.492,
    "end": 1621.077,
    "text": "- You talked about\nthese recursive loops."
  },
  {
    "start": 1621.077,
    "end": 1622.954,
    "text": "I think, that even in the past release,"
  },
  {
    "start": 1622.954,
    "end": 1624.998,
    "text": "we're starting to see\nthe very first times where"
  },
  {
    "start": 1624.998,
    "end": 1628.543,
    "text": "the embodied reasoning model can\nactually watch the robot do something,"
  },
  {
    "start": 1628.543,
    "end": 1630.086,
    "text": "and have opinions about it."
  },
  {
    "start": 1630.086,
    "end": 1630.629,
    "text": "And over time,"
  },
  {
    "start": 1630.629,
    "end": 1631.88,
    "text": "that actually starts\nto form a real loop."
  },
  {
    "start": 1631.88,
    "end": 1634.341,
    "text": "You're like, I think you should go\ncollect a little bit of different data."
  },
  {
    "start": 1634.341,
    "end": 1637.969,
    "text": "Or I think you should actually,\nand increasingly the researchers are asking,"
  },
  {
    "start": 1637.969,
    "end": 1640.93,
    "text": "\"Can you please provide an interface\nby which the higher level model"
  },
  {
    "start": 1640.93,
    "end": 1644.184,
    "text": "can actually give guiding\ninstructions to lower level models?\""
  },
  {
    "start": 1644.184,
    "end": 1645.894,
    "text": "It's actually painful\nto watch sometimes because"
  },
  {
    "start": 1645.894,
    "end": 1648.938,
    "text": "the high level model is like,\n\"No, no, just grab it a little higher.\""
  },
  {
    "start": 1648.938,
    "end": 1650.774,
    "text": "And so we're starting to\nsee these more and more."
  },
  {
    "start": 1650.774,
    "end": 1654.653,
    "text": "And so I do think, as more\nand more of the core capabilities"
  },
  {
    "start": 1654.653,
    "end": 1658.031,
    "text": "of Gemini spatial reasoning,\nand things we do with embodied reasoning"
  },
  {
    "start": 1658.031,
    "end": 1660.325,
    "text": "start to really get closer to AGI,"
  },
  {
    "start": 1660.325,
    "end": 1662.661,
    "text": "you will get some of those\nfeedback loops to start to form."
  },
  {
    "start": 1662.661,
    "end": 1664.079,
    "text": "- Yeah, that's super interesting."
  },
  {
    "start": 1664.079,
    "end": 1669.084,
    "text": "Well, let's look at a demo maybe,\nof the dexterous hands,"
  },
  {
    "start": 1669.084,
    "end": 1671.419,
    "text": "because I want to see,\nI want to see it come to life."
  },
  {
    "start": 1671.419,
    "end": 1672.045,
    "text": "- Sure. Yeah."
  },
  {
    "start": 1672.045,
    "end": 1675.256,
    "text": "So, here we're looking at GR2,\nand it's controlling"
  },
  {
    "start": 1675.256,
    "end": 1678.051,
    "text": "these very high degree of freedom hands."
  },
  {
    "start": 1678.051,
    "end": 1681.68,
    "text": "I think there are twenty different\njoints that it can control per hand."
  },
  {
    "start": 1681.68,
    "end": 1684.14,
    "text": "And the cool thing is\nthat we trained GR to control"
  },
  {
    "start": 1684.14,
    "end": 1686.559,
    "text": "the whole body and the\nhands with the same recipe."
  },
  {
    "start": 1686.559,
    "end": 1688.019,
    "text": "So, there's nothing\nspecial about the hand."
  },
  {
    "start": 1688.019,
    "end": 1690.647,
    "text": "It's just like we\ncollected much more diverse,"
  },
  {
    "start": 1690.647,
    "end": 1694.067,
    "text": "rich dexterous data,\nand the model is able to perform the tasks."
  },
  {
    "start": 1694.067,
    "end": 1695.568,
    "text": "So let's take a look\nat some of these tasks."
  },
  {
    "start": 1699.531,
    "end": 1700.907,
    "text": "- In order to be useful,"
  },
  {
    "start": 1700.907,
    "end": 1703.952,
    "text": "a robot needs the dexterity\nthat we take for granted."
  },
  {
    "start": 1703.952,
    "end": 1706.538,
    "text": "- You probably don't think about\nhow to drive twenty-two separate joints"
  },
  {
    "start": 1706.538,
    "end": 1711.251,
    "text": "when you operate your hand, but that's\nwhat we're asking these AI models to do."
  },
  {
    "start": 1711.251,
    "end": 1712.627,
    "text": "In the Gemini Robotics 2,"
  },
  {
    "start": 1712.627,
    "end": 1717.298,
    "text": "we came up with a set of tasks in\norder to test and develop dexterity."
  },
  {
    "start": 1723.722,
    "end": 1726.474,
    "text": "So we're asking the robot to pack lunch"
  },
  {
    "start": 1726.474,
    "end": 1729.185,
    "text": "by putting the grapes\ninto the Ziploc bag."
  },
  {
    "start": 1729.185,
    "end": 1734.19,
    "text": "So this requires a lot of precision,\nbut also a lot of coordination."
  },
  {
    "start": 1734.19,
    "end": 1738.528,
    "text": "Now, the really hard part\nis getting the Ziploc closed."
  },
  {
    "start": 1738.528,
    "end": 1740.864,
    "text": "Very nice job, Apollo."
  },
  {
    "start": 1740.864,
    "end": 1743.575,
    "text": "Hey Apollo, can you unscrew the bulb?"
  },
  {
    "start": 1743.575,
    "end": 1745.618,
    "text": "The bulb is actually a sphere, right?"
  },
  {
    "start": 1745.618,
    "end": 1748.413,
    "text": "So those contacts\nneed to be very precise."
  },
  {
    "start": 1748.413,
    "end": 1750.206,
    "text": "For it to be engaged\nwith the fingertips."
  },
  {
    "start": 1750.206,
    "end": 1751.624,
    "text": "There's actually a lot of motions"
  },
  {
    "start": 1751.624,
    "end": 1754.419,
    "text": "and a lot of dexterity\nthat's involved when you do that."
  },
  {
    "start": 1754.419,
    "end": 1756.963,
    "text": "You have to...\n- That was the easy Ziploc bag, too."
  },
  {
    "start": 1756.963,
    "end": 1758.548,
    "text": "I'm like, I can't even\ndo the regular ones."
  },
  {
    "start": 1758.548,
    "end": 1760.925,
    "text": "- that's actually very complicated."
  },
  {
    "start": 1760.925,
    "end": 1762.135,
    "text": "Thank you, Apollo."
  },
  {
    "start": 1762.135,
    "end": 1765.138,
    "text": "We're advancing what we can\ndo even with parallel grippers."
  },
  {
    "start": 1765.138,
    "end": 1769.851,
    "text": "Need to have dexterity,\nprecision, and 3D space understanding."
  },
  {
    "start": 1769.851,
    "end": 1771.853,
    "text": "- This is the robot\nthat you have behind you."
  },
  {
    "start": 1771.853,
    "end": 1774.105,
    "text": "- ...that we are trying to solve."
  },
  {
    "start": 1774.105,
    "end": 1777.233,
    "text": "We can move the kit around,\nwe can move the tools around."
  },
  {
    "start": 1777.233,
    "end": 1781.988,
    "text": "The robot is going to be able\nto understand how to reorient"
  },
  {
    "start": 1781.988,
    "end": 1786.91,
    "text": "the objects in space and\nthen precisely put them in."
  },
  {
    "start": 1786.91,
    "end": 1791.706,
    "text": "The robot's task is to tie a knot,\ntie off the trash bag."
  },
  {
    "start": 1791.706,
    "end": 1794.667,
    "text": "Multifingered hands are a key ingredient"
  },
  {
    "start": 1794.667,
    "end": 1798.922,
    "text": "for this kind of\nintricate knot-tying dexterity."
  },
  {
    "start": 1798.922,
    "end": 1801.508,
    "text": "Come on, robot, you got this."
  },
  {
    "start": 1801.508,
    "end": 1803.468,
    "text": "- I still have not\nfigured out how to do it."
  },
  {
    "start": 1803.468,
    "end": 1807.43,
    "text": "- Yeah, what?\nI've never seen a tie like that before."
  },
  {
    "start": 1807.43,
    "end": 1811.142,
    "text": "- We are pushing our\nunderstanding of how robots may interact"
  },
  {
    "start": 1811.142,
    "end": 1816.773,
    "text": "with complex objects in the real world,\ntrash bags, or hazardous waste."
  },
  {
    "start": 1816.773,
    "end": 1819.359,
    "text": "It would be great if we\ncould send a robot to do that,"
  },
  {
    "start": 1819.359,
    "end": 1822.487,
    "text": "rather than have humans\nput themselves at risk."
  },
  {
    "start": 1829.702,
    "end": 1830.995,
    "text": "- Very cool."
  },
  {
    "start": 1830.995,
    "end": 1837.001,
    "text": "And so in this example, the—\nthis is like a partner hardware"
  },
  {
    "start": 1837.001,
    "end": 1841.673,
    "text": "that we've generalized\nthe ER model to be able to"
  },
  {
    "start": 1841.673,
    "end": 1845.468,
    "text": "work on that specific set\nof hands in that context?"
  },
  {
    "start": 1845.468,
    "end": 1848.638,
    "text": "- So, this is the Gemini,\nthe VLA, the action model."
  },
  {
    "start": 1848.638,
    "end": 1849.43,
    "text": "- The action model. Okay.\n- Yeah."
  },
  {
    "start": 1849.43,
    "end": 1854.102,
    "text": "So, that one is trained to then\nuse these high dexterity hands."
  },
  {
    "start": 1854.102,
    "end": 1858.147,
    "text": "So, we did collect teleoperation\ndata to see how the task can be done."
  },
  {
    "start": 1858.147,
    "end": 1861.192,
    "text": "And then that data helps the model\nunderstand how to control these robots."
  },
  {
    "start": 1861.192,
    "end": 1867.49,
    "text": "- And does the dexterous hand use case generalize?\nOr is that also something"
  },
  {
    "start": 1867.49,
    "end": 1870.869,
    "text": "that as you see across all the\ndifferent hardware robotic partners,"
  },
  {
    "start": 1870.869,
    "end": 1873.037,
    "text": "the hands are all different\nor the degrees of freedom"
  },
  {
    "start": 1873.037,
    "end": 1875.373,
    "text": "are different and that's\nwhy it makes it complicated?"
  },
  {
    "start": 1875.373,
    "end": 1876.457,
    "text": "- So, I think that hands are a good"
  },
  {
    "start": 1876.457,
    "end": 1878.543,
    "text": "place to just push\nthe limits of dexterity,"
  },
  {
    "start": 1878.543,
    "end": 1881.588,
    "text": "but we are seeing some really cool\nsigns of cross-embodiment transfer."
  },
  {
    "start": 1881.588,
    "end": 1885.216,
    "text": "In our GR 1.5 release,\nwe talked about this more explicitly where, yeah,"
  },
  {
    "start": 1885.216,
    "end": 1888.595,
    "text": "we are seeing transfer between\nthe gripper task and the hands task."
  },
  {
    "start": 1888.595,
    "end": 1891.222,
    "text": "So, these models,\nwhen we train it with all the data,"
  },
  {
    "start": 1891.222,
    "end": 1892.974,
    "text": "we don't train per-embodiment models."
  },
  {
    "start": 1892.974,
    "end": 1895.602,
    "text": "GR 2 is trained on many robots."
  },
  {
    "start": 1895.602,
    "end": 1898.438,
    "text": "And we do see these signs of life,\nlike it understands basic concepts,"
  },
  {
    "start": 1898.438,
    "end": 1902.025,
    "text": "and it can transfer that action\nfrom one robot to the other."
  },
  {
    "start": 1902.025,
    "end": 1904.444,
    "text": "- And this is actually\nreally important because I think"
  },
  {
    "start": 1904.444,
    "end": 1906.696,
    "text": "robots will continue\nto evolve all the time."
  },
  {
    "start": 1906.696,
    "end": 1910.992,
    "text": "And we see it even in\nall the robots that we have every year,"
  },
  {
    "start": 1910.992,
    "end": 1913.077,
    "text": "like they evolve in\nsome interesting way, right?"
  },
  {
    "start": 1913.077,
    "end": 1915.58,
    "text": "And the hands is one\nof those areas that is"
  },
  {
    "start": 1915.58,
    "end": 1919.042,
    "text": "very ripe for a lot of\nacceleration over the next year."
  },
  {
    "start": 1919.042,
    "end": 1922.17,
    "text": "So, we fully expect the\nhands to be changing constantly."
  },
  {
    "start": 1922.17,
    "end": 1925.381,
    "text": "So, I think it is really\nimportant to enable models"
  },
  {
    "start": 1925.381,
    "end": 1927.508,
    "text": "that can work across all\nthese different embodiments."
  },
  {
    "start": 1927.508,
    "end": 1931.304,
    "text": "And fundamentally,\nif you're actually doing a task"
  },
  {
    "start": 1931.304,
    "end": 1933.473,
    "text": "where you're organizing things,"
  },
  {
    "start": 1933.473,
    "end": 1936.768,
    "text": "I mean, 90% of that task is not\nabout exactly how you move your hands."
  },
  {
    "start": 1936.768,
    "end": 1939.646,
    "text": "It's about understanding\nwhere you're putting things."
  },
  {
    "start": 1939.646,
    "end": 1943.191,
    "text": "And then the last 10% is about exactly\nhow you move your hand to achieve it."
  },
  {
    "start": 1943.191,
    "end": 1945.485,
    "text": "And so a lot of that\ntransfers between robots."
  },
  {
    "start": 1945.485,
    "end": 1946.945,
    "text": "Of course, there is some limitations,"
  },
  {
    "start": 1946.945,
    "end": 1949.948,
    "text": "like a gripper can\nonly grasp things this way,"
  },
  {
    "start": 1949.948,
    "end": 1953.868,
    "text": "and a hand could actually\ndo something more complex."
  },
  {
    "start": 1953.868,
    "end": 1957.83,
    "text": "But there's a lot of semantics\nthat are shared between them."
  },
  {
    "start": 1957.83,
    "end": 1962.335,
    "text": "- I'm curious actually if\nmodels having code quality"
  },
  {
    "start": 1962.335,
    "end": 1965.463,
    "text": "is at all correlated\nwith some of these use cases."
  },
  {
    "start": 1965.463,
    "end": 1971.052,
    "text": "And maybe my mental model is off on this,\nbut you imagine you can"
  },
  {
    "start": 1971.052,
    "end": 1973.972,
    "text": "deterministically program\nrobots in certain cases."
  },
  {
    "start": 1973.972,
    "end": 1977.183,
    "text": "And so could you,\nI'm curious if we do anything around that,"
  },
  {
    "start": 1977.183,
    "end": 1981.062,
    "text": "or if that's actually a use case that's helped,\nassuming we get"
  },
  {
    "start": 1981.062,
    "end": 1983.69,
    "text": "super intelligence at code,\nwhatever, in the next couple of years"
  },
  {
    "start": 1983.69,
    "end": 1987.36,
    "text": "because we're hill climbing it,\ndoes that somehow help?"
  },
  {
    "start": 1987.36,
    "end": 1989.862,
    "text": "You could almost\ndeterministically program"
  },
  {
    "start": 1989.862,
    "end": 1994.117,
    "text": "sequences of things that the robots are doing,\nor is that not?"
  },
  {
    "start": 1994.117,
    "end": 1995.827,
    "text": "- I can give you one example."
  },
  {
    "start": 1995.827,
    "end": 1999.247,
    "text": "- So, I think one place that can\nreally come into play is in simulation."
  },
  {
    "start": 1999.247,
    "end": 2002.75,
    "text": "So on the real robot,\nit's very difficult to write deterministic code"
  },
  {
    "start": 2002.75,
    "end": 2004.919,
    "text": "that can actually take\nin just a raw set of pixels"
  },
  {
    "start": 2004.919,
    "end": 2005.837,
    "text": "from a bunch of different cameras,"
  },
  {
    "start": 2005.837,
    "end": 2007.755,
    "text": "and actually give you thoughtful,\ncorrect joint angles."
  },
  {
    "start": 2007.755,
    "end": 2009.507,
    "text": "So, you can play some\ngames with inverse kinematics,"
  },
  {
    "start": 2009.507,
    "end": 2010.717,
    "text": "but it's really hard."
  },
  {
    "start": 2010.717,
    "end": 2011.342,
    "text": "In the simulator,"
  },
  {
    "start": 2011.342,
    "end": 2013.261,
    "text": "you often have access\nto privileged information."
  },
  {
    "start": 2013.261,
    "end": 2015.68,
    "text": "So, you actually secretly\nknow exactly how far away"
  },
  {
    "start": 2015.68,
    "end": 2017.515,
    "text": "this lid is from my fingers."
  },
  {
    "start": 2017.515,
    "end": 2019.434,
    "text": "And so if you can get to a\npoint where you're really starting to"
  },
  {
    "start": 2019.434,
    "end": 2023.062,
    "text": "build confidence, where the simulator\nis actually either a source of data,"
  },
  {
    "start": 2023.062,
    "end": 2025.064,
    "text": "or a place you want\nto evaluate a policy,"
  },
  {
    "start": 2025.064,
    "end": 2027.65,
    "text": "now, you can basically\nuse that code to guide"
  },
  {
    "start": 2027.65,
    "end": 2031.946,
    "text": "yourself in much more precisely,\nbecause you actually have access to"
  },
  {
    "start": 2031.946,
    "end": 2036.367,
    "text": "really correct resolution,\nreally precise information, rather than forcing"
  },
  {
    "start": 2036.367,
    "end": 2039.328,
    "text": "that code to kind of interpret sort\nof the messiness of the real world."
  },
  {
    "start": 2039.328,
    "end": 2040.663,
    "text": "- Yeah, yeah, yeah.\nNo, that makes sense."
  },
  {
    "start": 2040.663,
    "end": 2041.456,
    "text": "- Interesting."
  },
  {
    "start": 2041.456,
    "end": 2044.042,
    "text": "- Even for accelerating\nthe research loop, right?"
  },
  {
    "start": 2044.042,
    "end": 2045.501,
    "text": "We're using agents today, right?"
  },
  {
    "start": 2045.501,
    "end": 2049.005,
    "text": "To be able to run experiments,\nsee what works, see what did not work,"
  },
  {
    "start": 2049.005,
    "end": 2050.089,
    "text": "plot the differences,"
  },
  {
    "start": 2050.089,
    "end": 2052.133,
    "text": "detect that something is\ngoing the wrong direction early,"
  },
  {
    "start": 2052.133,
    "end": 2054.927,
    "text": "and then change parameters.\nAll of that is already happening."
  },
  {
    "start": 2054.927,
    "end": 2056.804,
    "text": "Right. So, in that sense, yeah."
  },
  {
    "start": 2056.804,
    "end": 2059.599,
    "text": "- It's exciting.\n- So those will not be simple code."
  },
  {
    "start": 2059.599,
    "end": 2064.812,
    "text": "So I think that we have tried to write\ncode to control robots for decades."
  },
  {
    "start": 2064.812,
    "end": 2065.063,
    "text": "- Right."
  },
  {
    "start": 2065.063,
    "end": 2066.939,
    "text": "- So, because those code,\nif it's rule-based,"
  },
  {
    "start": 2066.939,
    "end": 2070.068,
    "text": "it's very hard to generalize\nto all kinds of environments."
  },
  {
    "start": 2070.068,
    "end": 2073.321,
    "text": "This is why we are switching to\nthis very data-driven paradigm."
  },
  {
    "start": 2073.321,
    "end": 2076.491,
    "text": "But in theory,\nall the neural networks, training,"
  },
  {
    "start": 2076.491,
    "end": 2079.744,
    "text": "data-driven, optimization,\nthey're still within the code space."
  },
  {
    "start": 2079.744,
    "end": 2082.038,
    "text": "We still write code to\ngenerate all these things."
  },
  {
    "start": 2082.038,
    "end": 2084.415,
    "text": "So, I think eventually it's possible,"
  },
  {
    "start": 2084.415,
    "end": 2088.169,
    "text": "but it requires a lot of auto\nresearch to make that happen."
  },
  {
    "start": 2088.169,
    "end": 2088.92,
    "text": "- Yeah."
  },
  {
    "start": 2088.92,
    "end": 2093.716,
    "text": "- I was thinking about\nroutines almost very explicit,"
  },
  {
    "start": 2093.716,
    "end": 2098.346,
    "text": "I guess, maybe the action space is\ntoo unconstrained to make that happen,"
  },
  {
    "start": 2098.346,
    "end": 2100.765,
    "text": "but could I, I don't know."
  },
  {
    "start": 2100.765,
    "end": 2104.185,
    "text": "I was thinking of examples where like it might,\nyou could do something useful,"
  },
  {
    "start": 2104.185,
    "end": 2108.106,
    "text": "if you could write code to do some\nof these use cases deterministically."
  },
  {
    "start": 2108.106,
    "end": 2112.11,
    "text": "But no, I hear what you're\nsaying that it doesn't generalize well."
  },
  {
    "start": 2112.11,
    "end": 2114.529,
    "text": "Well, so how can people actually\nstart getting access to the model?"
  },
  {
    "start": 2114.529,
    "end": 2116.364,
    "text": "I feel like it's,\nwe've made a bunch of progress."
  },
  {
    "start": 2116.364,
    "end": 2117.532,
    "text": "What's the availability story?"
  },
  {
    "start": 2117.532,
    "end": 2119.534,
    "text": "Where can people\nactually start using it?"
  },
  {
    "start": 2119.534,
    "end": 2120.576,
    "text": "- Yeah, we're really excited."
  },
  {
    "start": 2120.576,
    "end": 2124.163,
    "text": "So, Embodied Reasoning 2.0\nor 2 is going to come out,"
  },
  {
    "start": 2124.163,
    "end": 2126.332,
    "text": "that'll be directly\navailable via AI Studio."
  },
  {
    "start": 2126.332,
    "end": 2128.292,
    "text": "And I think I can say this\nbut soon it'll be available"
  },
  {
    "start": 2128.292,
    "end": 2131.254,
    "text": "via the Gemini Enterprise"
  },
  {
    "start": 2131.254,
    "end": 2134.423,
    "text": "- Agents Platform.\n- Agents Platform."
  },
  {
    "start": 2134.423,
    "end": 2139.303,
    "text": "- But Gemini Robotics 2,\nwe'll have the embodied reasoning model"
  },
  {
    "start": 2139.303,
    "end": 2142.515,
    "text": "available directly via an API,\nand that will be generally accessible,"
  },
  {
    "start": 2142.515,
    "end": 2144.517,
    "text": "and we really\nencourage people to use that."
  },
  {
    "start": 2144.517,
    "end": 2147.895,
    "text": "And then the action models themselves\nare also going to be available."
  },
  {
    "start": 2147.895,
    "end": 2149.397,
    "text": "We work directly with our deep partners,"
  },
  {
    "start": 2149.397,
    "end": 2152.942,
    "text": "so they'll be the ones who use\nthe biggest strongest versions."
  },
  {
    "start": 2152.942,
    "end": 2156.32,
    "text": "We also have an on-device version\nof that, that our trusted testers can use."
  },
  {
    "start": 2156.32,
    "end": 2158.656,
    "text": "And so people are welcome to join."
  },
  {
    "start": 2158.656,
    "end": 2160.825,
    "text": "We currently have a waitlist-- but\nwe're trying to do more about it--"
  },
  {
    "start": 2160.825,
    "end": 2162.41,
    "text": "our trusted tester program."
  },
  {
    "start": 2162.41,
    "end": 2165.746,
    "text": "And then, they can actually get access\nto an on-device deployable version"
  },
  {
    "start": 2165.746,
    "end": 2166.831,
    "text": "of the action model, where they can"
  },
  {
    "start": 2166.831,
    "end": 2169.167,
    "text": "actually fine-tune that\nmodel directly on either"
  },
  {
    "start": 2169.167,
    "end": 2172.044,
    "text": "their tasks or their robots and\nactually try it out in practice."
  },
  {
    "start": 2172.044,
    "end": 2173.337,
    "text": "- Yeah, that's awesome. I'm excited."
  },
  {
    "start": 2173.337,
    "end": 2174.922,
    "text": "What--"
  },
  {
    "start": 2174.922,
    "end": 2178.509,
    "text": "any advice? And maybe maybe I'm\nmisremembering this, but I feel like we"
  },
  {
    "start": 2178.509,
    "end": 2182.305,
    "text": "do have a bunch of customers who\nuse the ER models who actually aren’t"
  },
  {
    "start": 2182.305,
    "end": 2183.598,
    "text": "robotics companies, and they just"
  },
  {
    "start": 2183.598,
    "end": 2186.309,
    "text": "happen to be in one of\nthese domains for video,"
  },
  {
    "start": 2186.309,
    "end": 2188.728,
    "text": "audio, spatial understanding,\nor something like that."
  },
  {
    "start": 2188.728,
    "end": 2192.023,
    "text": "I don't know if that's a\nsuggested path for folks,"
  },
  {
    "start": 2192.023,
    "end": 2196.903,
    "text": "but I assume it's those domains of video,\nspatial understanding,"
  },
  {
    "start": 2196.903,
    "end": 2200.364,
    "text": "where the ER model is better on\na bunch of these core benchmarks."
  },
  {
    "start": 2200.364,
    "end": 2202.7,
    "text": "- Yeah, I mean,\nthese models are better in a few ways."
  },
  {
    "start": 2202.7,
    "end": 2204.452,
    "text": "The ER model in particular."
  },
  {
    "start": 2204.452,
    "end": 2206.954,
    "text": "This release is a lot better\nat video understanding."
  },
  {
    "start": 2206.954,
    "end": 2209.916,
    "text": "So before, it's always been very good,"
  },
  {
    "start": 2209.916,
    "end": 2211.918,
    "text": "and state-of-the-art\nat spatial understanding."
  },
  {
    "start": 2211.918,
    "end": 2214.17,
    "text": "So 2D and 3D bounding box,"
  },
  {
    "start": 2214.17,
    "end": 2216.422,
    "text": "understanding where\nobjects are in 3D space."
  },
  {
    "start": 2216.422,
    "end": 2217.882,
    "text": "Now, it can understand videos."
  },
  {
    "start": 2217.882,
    "end": 2220.801,
    "text": "It understands also\nthe semantics of a task."
  },
  {
    "start": 2220.801,
    "end": 2221.677,
    "text": "So if you ask it,"
  },
  {
    "start": 2221.677,
    "end": 2223.721,
    "text": "at what point should I\nstop pouring my coffee?"
  },
  {
    "start": 2223.721,
    "end": 2226.307,
    "text": "Or am I done closing this Ziploc bag?"
  },
  {
    "start": 2226.307,
    "end": 2229.227,
    "text": "It actually understands how far\nalong you are in that progress."
  },
  {
    "start": 2229.227,
    "end": 2231.562,
    "text": "And so that's extremely\nuseful whether you're doing"
  },
  {
    "start": 2231.562,
    "end": 2235.775,
    "text": "any kind of like video understanding capability,\nor for robotics, exactly."
  },
  {
    "start": 2235.775,
    "end": 2238.653,
    "text": "If you use it as your agent,\nthen now, it can be the agent"
  },
  {
    "start": 2238.653,
    "end": 2241.322,
    "text": "that understands how far along you are,\nand decides to"
  },
  {
    "start": 2241.322,
    "end": 2244.742,
    "text": "switch to a different task\nor decide that you're done."
  },
  {
    "start": 2244.742,
    "end": 2247.62,
    "text": "And then these models are\nalso significantly safer."
  },
  {
    "start": 2247.62,
    "end": 2253.709,
    "text": "This is our safest model yet,\nand it's safer not only in the regular way"
  },
  {
    "start": 2253.709,
    "end": 2257.421,
    "text": "in which all of our Gemini models are safe,\nin terms of content safety,"
  },
  {
    "start": 2257.421,
    "end": 2261.008,
    "text": "but it's also safer\nbecause it understands"
  },
  {
    "start": 2261.008,
    "end": 2266.013,
    "text": "the likelihood that a model is\ngoing to be completing a task."
  },
  {
    "start": 2266.013,
    "end": 2269.85,
    "text": "So, it also helps you\nunderstand the— if you give"
  },
  {
    "start": 2269.85,
    "end": 2272.561,
    "text": "an instruction that is very ambiguous,\nfor example, then,"
  },
  {
    "start": 2272.561,
    "end": 2276.274,
    "text": "it will ask proactively the human,\n\"Oh, that instruction is very ambiguous."
  },
  {
    "start": 2276.274,
    "end": 2277.525,
    "text": "What do you mean?\""
  },
  {
    "start": 2277.525,
    "end": 2279.819,
    "text": "And so it helps with\nproactive clarification."
  },
  {
    "start": 2279.819,
    "end": 2281.279,
    "text": "And then the last one is that we're also"
  },
  {
    "start": 2281.279,
    "end": 2283.906,
    "text": "making it really\nstrong at detecting humans,"
  },
  {
    "start": 2283.906,
    "end": 2286.701,
    "text": "and humans' proximity to robots,\nwhich is very important when"
  },
  {
    "start": 2286.701,
    "end": 2290.288,
    "text": "you're talking about collaborative\nrobots that are in human-centric spaces."
  },
  {
    "start": 2290.288,
    "end": 2291.998,
    "text": "So those are just a few areas."
  },
  {
    "start": 2291.998,
    "end": 2295.793,
    "text": "We're also introducing a new safety\nbenchmark that we have open source."
  },
  {
    "start": 2295.793,
    "end": 2298.546,
    "text": "It's called Asimov Agentic."
  },
  {
    "start": 2298.546,
    "end": 2304.719,
    "text": "Asimov is our benchmark and it\nessentially has a large set of examples,"
  },
  {
    "start": 2304.719,
    "end": 2309.223,
    "text": "real-world examples, where you have\nto make a decision about what the robot"
  },
  {
    "start": 2309.223,
    "end": 2313.644,
    "text": "would do next or what you should\ndo next based on this situation."
  },
  {
    "start": 2313.644,
    "end": 2316.439,
    "text": "So it's a lot about\nsemantic physical understanding."
  },
  {
    "start": 2316.439,
    "end": 2319.15,
    "text": "It's a lot about common sense\nthat robots would need to have,"
  },
  {
    "start": 2319.15,
    "end": 2322.945,
    "text": "if they're going to be operating\nand doing lots of tasks around us."
  },
  {
    "start": 2322.945,
    "end": 2325.031,
    "text": "- Very cool.\nThis was an awesome conversation."
  },
  {
    "start": 2325.031,
    "end": 2326.949,
    "text": "It was super interesting\nto hear about the launches."
  },
  {
    "start": 2326.949,
    "end": 2329.41,
    "text": "I'm very excited for folks to\nget their hands on the models."
  },
  {
    "start": 2329.41,
    "end": 2331.662,
    "text": "It's cool to also come to y'all's space."
  },
  {
    "start": 2331.662,
    "end": 2334.54,
    "text": "I feel like there's,\nit's very-- much more interesting"
  },
  {
    "start": 2334.54,
    "end": 2336.167,
    "text": "than the normal Google offices."
  },
  {
    "start": 2336.167,
    "end": 2339.128,
    "text": "So, I'm glad to be a guest and see"
  },
  {
    "start": 2339.128,
    "end": 2340.546,
    "text": "all the cool hard work\nthat you all are doing."
  },
  {
    "start": 2340.546,
    "end": 2342.048,
    "text": "So congrats on the launch."
  },
  {
    "start": 2342.048,
    "end": 2345.259,
    "text": "Very excited. Thanks everyone for\nwatching this episode of Release Notes."
  },
  {
    "start": 2345.259,
    "end": 2346.302,
    "text": "We'll see you in the next one."
  }
]