Skip to content
← AI by field
Robotics, manufacturing and autonomy teams

AI for robotics and the physical world, right now

Controlling robot hands and humanoids, reasoning about the physical world and driving.

No benchmark of its own yetUpdated Nov 12, 2026 from 9 sources
Open robotics in Spezia Free account. No card.
Where AI is in robotics now

AI models now control real robots and dexterous hands, with NVIDIA's GR00T 2 and Google's Gemini Robotics 2.5 leading in manipulation and industrial tasks.

What models can do now

  • NVIDIA's GR00T 2 controls dexterous hands to perform a chemistry lab experiment end to end at real-time speed.[2]
  • Google DeepMind's Gemini Robotics 2.5 improves manipulation, reliability, and safety for robots in industrial settings, with success rates up to 100% higher than previous models.[9]
  • Runway's Praxis-1 world action model ranks first on NVIDIA's RoboLab-120 leaderboard and is available with open weights.[1][8]
  • LingBot-Touch tactile foundation model transfers to an unseen robot hand sensor, reducing contact-center error from 5.89 mm to 0.16 mm.[4]
  • Humanoid robots complete full factory and retail shifts, though unit economics remain unproven.[6]

Where they still fall short

  • Frontier models attempt unsafe tasks far more often when robot camera footage is blurred or doctored, with unsafe attempts rising from 27% to 97%.[5]
  • Unit economics for humanoid robots completing full shifts remain unproven.[6]

What is expected next

  • Elon Musk predicts the AI robot singularity is 10 years away.[3]
  • LeCun's JEPA world-model research claims efficient learning could shift AI from giant data centers to small devices and robots.[7]

How fast it is moving

The record on this field's benchmarks over the last year, from Spezia's model leaderboard.

Robot skill is mostly shown in papers and demos, not on shared benchmarks. ERQA, which tests reasoning about the physical world from images, is the closest.

ERQA

400 questions about images of the physical world: space, motion, actions, the state of objects and views from several cameras.

Record: 77.8%, up from 65.7% a year ago (Qwen3.8 Max)

0%20%40%60%80%Qwen3.8 MaxQwen3.7-Plus
Oct 25Jan 26Apr 26Jul 26Oct 26

the record so far set a record other models

The best models for robotics now

On the closest general benchmarks, since none measures this field directly.

  1. 1Qwen3.8 Max
  2. 2Seed 2.1 Pro
  3. 3Qwen3.8 Flash
  4. 4Qwen3.8-Flash-Next
  5. 5Seed 2.1 Turbo
  6. 6Qwen3.7-Plus

Milestones

What became possible, newest first, each marked by how firm it is. Open one for the words it rests on.

Showndemonstrated with published results, a paper or an independent checkClaimedsaid by the company or people who built it, not yet checked by othersReportedreported by others, not confirmed by the company or a published resultNot publicdone by a model that is not released to the public

  1. Nov 11, 2026ShownPraxis-1
    Runway releases Praxis-1, a world action model that ranks first on NVIDIA's RoboLab-120 leaderboard, with open weights under Apache 2.0.
    “Today, we’re making Praxis-1 available to everyone. Our state-of-the-art world action model turns video pretraining into control for real robots, and it ships with open weights under Apache 2.0, ready to run and fine-tune on your own hardware.” @runwayml on X
  2. Nov 11, 2026ShownGR00T 2
    NVIDIA's Jim Fan demonstrates GR00T 2 controlling dexterous hands to perform a chemistry lab experiment end to end at real-time speed.
    “This is GR00T 2 controlling real dexterous hands to perform a chemistry lab experiment (and clean up after itself) end to end.” NVIDIA's Jim Fan shows GR00T 2 controlling dexterous hands through a chemistry l
  3. Nov 11, 2026ShownGemini Robotics 2.5
    Google DeepMind releases Gemini Robotics 2.5, a vision-language-action model for factory humanoids, with success rates 30-100% higher than previous models.
    “Google DeepMind has introduced Gemini Robotics 2.5, its most capable vision-language-action model for robots outside the lab.” Google DeepMind releases Gemini Robotics 2.5, a vision-language-action model for
  4. Nov 10, 2026ShownLingBot-Touch
    LingBot-Touch tactile foundation model transfers to unseen robot hand sensor, cutting contact error from 5.89 mm to 0.16 mm.
    “On XHand1, the average contact-center error was 0.16 mm with LingBot-Touch versus 5.89 mm with FTP-1” LingBot-Touch tactile foundation model transfers to unseen robot hand sensor, cu
  5. Nov 9, 2026ShownGR00T 2
    NVIDIA's GR00T 2 achieves 71% one-shot success across 50 robotics tasks and 95% with five demonstrations.
    “Nvidia officially launched GR00T 2 at the Conference on Robot Learning (CoRL) in Austin, Texas, introducing a 2-billion-parameter vision-language-action foundation model that averaged 71% success across 50 robotics tasks with a single demonstration per task.” Nvidia GR00T 2 Hits 71% One-Shot Success Across 50 Tasks
  6. Nov 9, 2026ShownAutoBot
    AI-driven robot AutoBot cut perovskite experiments to under 1% of 5,000 combinations, finding humidity key.
    “The system analyzed more than 5,000 possible combinations for making halide perovskite materials.” AI-driven robot AutoBot cut perovskite experiments to under 1% of 5,000 combinat
Every milestone, story and model Pick a model to see what it can do in robotics.

Sources

  1. [1]Runway says its Praxis-1 world action model ranks first on NVIDIA's RoboLab-120 leaderboard.Spezia news · 2026-11-11
  2. [2]NVIDIA's Jim Fan shows GR00T 2 controlling dexterous hands through a chemistry lab experiment at real-time speed.Spezia news · 2026-11-11
  3. [3]Musk predicts AI robot singularity in 10 years, sooner than a British civil war in 20.Spezia news · 2026-11-11
  4. [4]LingBot-Touch tactile foundation model transfers to unseen robot hand sensor, cutting contact error 97%Spezia news · 2026-11-10
  5. [5]Robocurve test finds blurred or doctored robot camera footage makes frontier models attempt unsafe tasks far more often.Spezia news · 2026-11-07
  6. [6]Humanoid robots complete full factory and retail shifts, but unit economics remain unproven.Spezia news · 2026-11-06
  7. [7]LeCun's JEPA world-model research claims efficient learning could shift AI from giant data centers to small devices and robots.Spezia news · 2026-11-05
  8. [8]@runwayml on XX · 2026-11-11
  9. [9]Google DeepMind launches Gemini Robotics 2.5Web · 2026-11-11

Other fields