{
  "headline": "The Last Millimeter: Robots Are Learning to Feel Their Way Through Real Work",
  "slug": "the-last-millimeter",
  "url": "https://emberandoakinstitute.com/blog/the-last-millimeter.html",
  "date": "2026-10-20",
  "publisher": "Ember & Oak Institute, LLC",
  "category": "AI news",
  "tags": [
    "robotics",
    "physical AI",
    "robot foundation models",
    "dexterous manipulation",
    "proprioception",
    "humanoid robots",
    "AI news"
  ],
  "summary": "A new robot brain ran 27 robots it had never seen. A startup's force-sensing arm is wowing veteran roboticists. MIT researchers taught a robot to fit parts together with no camera at all, and a humanoid has walked for nine days straight. This week's robot news says the hard part is no longer knowing what to do. It is the touch, the fit and the stamina to do it for real.",
  "key_facts": [
    {
      "id": 1,
      "claim": "Dream Labs came out of stealth on Oct. 19, 2026 with Dream-1, a World Action Model it says controls bimanual parallel-gripper robots it has never seen, on unseen tasks, in unseen places, through a public API; beta testers ran it in 8 countries.",
      "sources": [
        "https://x.com/dreamlabs_ai/status/2112230668805828801",
        "https://dreamlabs.ai/research/dream-1/"
      ]
    },
    {
      "id": 2,
      "claim": "Dream Labs reports zero-shot control of 27 unseen robots in simulation, including the 2008-era Willow Garage PR2, with no embodiment-specific data; performance varies by robot and depends on each robot's inverse kinematics solver.",
      "sources": [
        "https://dreamlabs.ai/research/dream-1/"
      ]
    },
    {
      "id": 3,
      "claim": "Dream-1 predicts video and actions together and can output end-effector (gripper) poses, which each robot's IK solver converts into joint commands.",
      "sources": [
        "https://dreamlabs.ai/research/dream-1/"
      ]
    },
    {
      "id": 4,
      "claim": "Robocurve's independent RoboDojo-RC Tier 1 evaluation (report dated Oct. 18, 2026; 120 attempts per configuration, partial credit) scored Dream-1 59.3%, Gemini 4 Argon (high) 48.2%, GPT-6 Astra (high) 42.5%, GPT-6.1 Sol (high) 41.7%, Opus 5.5 (high) 38.2%, Gemini 3.8 Flash (high) 36.2%, MolmoAct2 18.9%, GR00T N1.7 9.5%.",
      "sources": [
        "https://robocurve.org/",
        "https://x.com/dreamlabs_ai/status/2112230800926331293"
      ]
    },
    {
      "id": 5,
      "claim": "RoboDojo-RC Tier 1 is run with two YAM arms, and YAM is the robot Dream-1 was trained on.",
      "sources": [
        "https://robocurve.org/rci/",
        "https://dreamlabs.ai/research/dream-1/"
      ]
    },
    {
      "id": 6,
      "claim": "Dream Labs calls Dream-1 a research beta that will make mistakes and has not been validated for safety-critical use.",
      "sources": [
        "https://dreamlabs.ai/research/dream-1/"
      ]
    },
    {
      "id": 7,
      "claim": "Eka Robotics, founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, announced a Series A on Oct. 20, 2026 led by a16z and F-Prime; it builds a vision-force-action model and a force-sensing, flexible gripper trained largely with simulation-based reinforcement learning.",
      "sources": [
        "https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/",
        "https://a16z.com/announcement/investing-in-eka-robotics/"
      ]
    },
    {
      "id": 8,
      "claim": "Eka's arm picks about 1,500 objects up to 1 kg per hour versus about 600 for a human, and Eka says it picks reliably more than 95% of the time (company figures reported by TechCrunch).",
      "sources": [
        "https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/"
      ]
    },
    {
      "id": 9,
      "claim": "Agrawal: 'Vision tells you roughly where the object is, what the pose is, but the actual grabbing, the actual searching is happening through forces.' Kris Hauser (UIUC) said it adapted to many objects without retraining and would try over and over if it failed, and that the big challenge is running on real assembly lines without breaking for an entire year.",
      "sources": [
        "https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/"
      ]
    },
    {
      "id": 10,
      "claim": "MIT's Bronars, Park, Kedia and Agrawal (arXiv:2610.23849, Oct. 16, 2026) trained blind assembly policies using only a 0.6 s history of joint positions, velocities and prior actions, entirely in simulation, for peg insertion, gear meshing and nut threading; zero-shot on a real robot the peg policy inserted eight peg shapes (three out of distribution) with 0.5-1.5 mm clearance per side in 73% of 237 trials; in simulation blind policies generalized far better than matched visual policies to unseen shapes and sizes, though they can be slower; the assembly generator is open source.",
      "sources": [
        "https://arxiv.org/abs/2610.23849",
        "https://arxiv.org/html/2610.23849v1",
        "https://younghyopark.me/blind-proprioceptive-assembly/"
      ]
    },
    {
      "id": 11,
      "claim": "On Oct. 19, 2026, Figure AI CEO Brett Adcock posted that Figure 03 had crossed 9 straight days walking on a treadmill, 348 miles and 1.2 million steps, started as a side project to stress the robot for failures.",
      "sources": [
        "https://x.com/adcock_brett/status/2112227583643509174"
      ]
    }
  ],
  "charts": [
    {
      "file": "https://emberandoakinstitute.com/blog/the-last-millimeter/robot-control-scores.png",
      "url": "https://emberandoakinstitute.com/blog/the-last-millimeter/assets/robot-control-scores.png",
      "alt": "Horizontal bar chart of mean scores on Robocurve's RoboDojo-RC Tier 1 robot test, run on two YAM arms: Dream-1 59.3%, Gemini 4 Argon high 48.2%, GPT-6 Astra high 42.5%, GPT-6.1 Sol high 41.7%, Opus 5.5 high 38.2%, Gemini 3.8 Flash high 36.2%, MolmoAct2 18.9%, GR00T N1.7 9.5%. Source: Robocurve, Oct. 2026.",
      "data": {
        "Dream-1": 59.3,
        "Gemini 4 Argon (high)": 48.2,
        "GPT-6 Astra (high)": 42.5,
        "GPT-6.1 Sol (high)": 41.7,
        "Opus 5.5 (high)": 38.2,
        "Gemini 3.8 Flash (high)": 36.2,
        "MolmoAct2": 18.9,
        "GR00T N1.7": 9.5
      },
      "unit": "mean score percent, RoboDojo-RC Tier 1",
      "source": "https://robocurve.org/"
    },
    {
      "file": "https://emberandoakinstitute.com/blog/the-last-millimeter/picks-per-hour.png",
      "url": "https://emberandoakinstitute.com/blog/the-last-millimeter/assets/picks-per-hour.png",
      "alt": "Bar chart comparing objects picked per hour, for items up to 1 kilogram: a typical person, about 600; Eka's robot arm, about 1,500, a company figure. Source: TechCrunch, Oct. 20, 2026.",
      "data": {
        "human_picks_per_hour": 600,
        "eka_picks_per_hour": 1500
      },
      "source": "https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/"
    }
  ],
  "hero": {
    "file": "https://emberandoakinstitute.com/blog/the-last-millimeter/hero.jpg",
    "url": "https://emberandoakinstitute.com/blog/the-last-millimeter/assets/hero.jpg",
    "alt": "Painterly storybook illustration of a cozy Mat-Su Valley log-cabin workshop on a snowy October morning: a parent and two young learners smile as a friendly wood-and-brass robot gently holds a raspberry in one padded hand and fits a wooden peg into a block with the other, while a tiny toy robot walks on a little treadmill in the background, a husky naps on a rag rug, and golden birches and mountains glow outside the window."
  }
}