FULL STORY

NVIDIA GR00T 2: From Teaser to Open-Source Release

After Jim Fan teased the reveal at CoRL, NVIDIA's GEAR team released and open-sourced GR00T 2, a world-action model running fully on-robot, with a humanoid demo autonomously performing chemistry experiments.

2026-11-09 ~ 2026-11-12 · 3 episodes · 25 posts

Episode 1 · NVIDIA to Unveil Gr00t 2 at CoRL Keynote (2026-11-09, 2 posts)

NVIDIA's Jim Fan teased the launch of Gr00t 2 during his CoRL keynote on November 9 in Austin, with the team also presenting four research works including GRAIL at the conference.

Episode 2 · NVIDIA Open-Sources GR00T 2, a World-Action Model for Robots (2026-11-09, 20 posts)

At CoRL 2026 in Austin, NVIDIA's GEAR team (led by Yuke Zhu and Jim Fan) released GR00T 2, a next-generation World-Action Model that simultaneously predicts future video and robot actions, letting the model "imagine" how the physical world unfolds before generating goal-reaching actions. The same model runs on parallel grippers, dexterous hands and humanoid robots. Jim Fan called it NVIDIA's most advanced open-source "System One" robot model: it runs entirely on onboard compute without tethering, weights are fully open (code opens in December), and it can chain 20+ steps of dexterous manipulation and whole-body control lasting over 5 minutes. Two related papers are finalists for CoRL Best Paper. Live demos covered bimanual fine manipulation (cutting, GPU installation, striking matches, slicing bananas, clearing tables) and whole-body mobility, with the model, data and simulation all being opened.

Confirmed

  • Performance: 40% zero-shot success rate across 50 tasks; 95% success on 5 demo tasks (per @CyberRobooo, training included 100k hours of egocentric human video)
  • Onboard running: the accelerated GR00T 2 Turbo runs in real time on Jetson Thor-equipped humanoids for walking, bending, reaching and carrying—demos included serving popcorn, wiping spills and clearing tables
  • Data Pyramid: 400k+ hours of egocentric human video, 10k hours of UMI handheld demos, and only 23.5k hours of robot teleoperation—most of the robot's knowledge comes from non-robot data
  • X-Adapt: skills demonstrated only by humans (bare hands, UMI grippers, wearable DexUMI) transfer directly to robots that never practiced them; a person wearing NVIDIA's wearable device made breakfast and the robot learned it with zero practice
  • WAM API: a reasoning model plans and checks task progress while GR00T 2 executes; agents can "point" in camera frames to instruct the robot. Per @shoumikgoswami, adding a reasoning agent lifted YAM robot toy-assembly success from 71-84 to 97-98 out of 100 runs, with recovery from failed steps; @markk added that each of three toys achieved at least 97/100 in consecutive tests via a closed loop of predicting expected camera frames and verifying after execution
  • Infrastructure (per @skypilotorg): a 2000-GPU cluster with highly mixed workloads, where pretraining accounts for only half the compute
  • Co-releases: Isaac Lab 3.0 and the Isaac Lab-Arena open industrial benchmark built with RLWRLD and UC Berkeley

Why it matters

  • World-action modeling—predicting how the physical world evolves before acting—is seen by @lukasmziegler as a fundamental capability boost and a path toward general robot intelligence
  • The Data Pyramid validates a scalable "human video first, robot data second" training paradigm, easing the robot-data scarcity bottleneck
  • X-Adapt and the WAM API lower the barriers to skill acquisition and task instruction; the reasoning agent's failure-recovery gains show that a "brain plans, body executes" split markedly improves long-horizon success rates
  • The rare disclosure of the 2000-GPU cluster's real workload mix offers a reference for understanding the compute needs of frontier robotics labs

Episode 3 · NVIDIA Open-Sources GR00T 2 as Robot Autonomously Runs Chemistry Lab (2026-11-11, 3 posts)

NVIDIA released GR00T 2, an open-source System One robot model that runs entirely on onboard compute, with demos showing a dexterous humanoid autonomously completing real chemistry experiments, including cleanup, in real time; observers say it appears close to production-ready.