NVIDIA GR00T 2: foundation world-action models for dexterous robots; whitepaper at CoRL, code in December

Why it matters

【Key points】

• NVIDIA Research GEAR released GR00T 2 at CoRL (Austin) on 2026-11-09: foundation world-action models (WAMs) that jointly predict future video and robot actions.
• Research page + whitepaper PDF live (S3 Last-Modified 13:34:26Z); companion blog 14:00:01Z.
• Page says Code Available in December — announce/paper today, not weights.
• Covers parallel grippers, dexterous hands, humanoids
• claims X-Adaptability, ~40% zero-shot / ~95% after 5 demos + 30-min RTX 5090 fine-tune on 50 unseen tasks, dual 22-DoF Sharpa dexterity (e.g. GB300 DIMM seating ~85%), agentic WAM API control, and on-robot GR00T 2 Turbo for Unitree G1.
• Human-1M-100K (100k hours) on Hugging Face.
• Blog also packages Isaac Lab 3.0 + industrial benchmarks — but Isaac Lab v3.0.0 GA was 2026-10-20; do not treat Isaac as new today.

【Our read】

Background

joint world+action prediction is the next step after action-only robot FMs; GR00T 2 puts cross-embodiment transfer and LLM/VLM-steered WAM control into cited numbers, with whitepaper-now / code-December CoRL timing.

Peers

• vs Dream-1 API and open VLA lines, this is NVIDIA full-stack (Cosmos/sim/Human-1M/Isaac) plus industrial contact tasks.
• Unlike Korea AI Day regional packs, dedicated research page + named model clears the frontier-primary bar; Isaac Lab 3.0 itself is not new.

Impact

raises expectations on human-video scale, cross-embodiment transfer, and agent scheduling; watch December code/license, real-robot replication, Turbo latency on G1. model_release (named WAM + whitepaper).

【Verifiable】

• research.nvidia.com/labs/gear/gr00t-2
• PDF.../gr00t2_whitepaper.pdf (LM 2026-11-09T13:34:26Z); blog(14:00:01Z); HF Human-1M-100K; IsaacLab v3.0.0 2026-10-20. Not on site; ≠ Flora/Dream-1. max e/1094. published_at null