Nvidia's GR00T 2 learns from plain human video: 40% zero-shot on 50 unseen robot tasks
rohanpaul_ai · x · 2026-11-11
Nvidia's GEAR lab released GR00T 2, a new generation of world-action models that predict future video and robot actions jointly, running on parallel grippers, dexterous hands, and humanoids. Code opens in December.
Highlights:
- Pretraining on ordinary human video: earlier models needed teleoperation data pairing frames with exact motor commands; GR00T 2's joint world+action prediction unlocks cheap, abundant command-free video.
- X-Adaptability: skills shown only by humans (lining a lab funnel, 65% success), a handheld UMI gripper (YAM arm assembling a toy dump truck end-to-end, 60%), or a wearable DexUMI hand (banana yogurt, striking a match) transfer to robots that never practiced them.
- Zero-/few-shot: 40% success zero-shot across 50 tasks unseen on the target robot; 5 demos (<3 min of robot data) per task sharply improve performance.
- Agentic control: because it predicts action outcomes, a reasoning model can direct it.
More from Embodied
- HIRO Launches Dual-Arm Robot Origin at $9,950 With $450 Refundable Deposit — Scobleizer · 2026-11-11
- Day 363 of Asimov: open-source humanoid tests new grippers for object manipulation — chris_j_paxton · 2026-11-11
- CoRL best-paper nominee Play2Perfect ported to new robots in IsaacLab and MJLab — leto__jean · 2026-11-11
- Walden Robotics lays out its embodied AI playbook: full-stack, sim, teleop, human data — chris_j_paxton · 2026-11-11
- One touch encoder generalizes across sensors, cutting contact error to 0.16mm — kimmonismus · 2026-11-11
- Chris Paxton rides a Zoox robotaxi for the first time since his internship 10 years ago — chris_j_paxton · 2026-11-11