MODEL
GR00T 2
Overview
GR00T 2 is NVIDIA’s second-generation general-purpose foundation model for humanoid robots, announced at the Conference on Robot Learning (CoRL) in Austin on 2026-11-09 and framed by NVIDIA as an “open frontier world-action model.” The model predicts how its actions will change what it sees before moving — the “world-action-model” framing is NVIDIA’s bet that robotics can be converted from a data-collection problem into a compute-scaling one. Training used 402,730 hours of egocentric human video, with GPT-6 Astra serving as the reasoning layer at inference time. NVIDIA reports 40% zero-shot success on 50 tasks absent from the target robot’s training data, and a toy-assembly aggregate of 292 of 300 successful trials across three toy types (~97%). The release pairs with Isaac Lab 3.0 as the simulated-data flywheel.
Timeline
- 2026-11-11-AI-Digest — NVIDIA launches GR00T 2 and Isaac Lab 3.0 at CoRL. Early-testing partners: Foxconn (factory automation for server manufacturing), Siemens (industrial use-case evaluation), Nexuni (robotic laundry). TNW explicitly notes none of the named partners has committed to production use yet; carried as “pilots, not factory-floor economics yet.” Separate DeepMind Gemini Robotics 2.5 “96% metal-parts sorting at Mercedes” figure referenced in some coverage could not be independently verified at the time of the digest.
Key Developments
-
World-Action Model Framing: NVIDIA positions GR00T 2 as predicting pixel-level consequences of actions before executing them — the explicit bet is that converting robotics into a compute-scaling problem beats the manual-data-collection regime that has defined prior humanoid work.
-
Industrial-Pilot Partner List: Foxconn, Siemens, Nexuni as named early-testing partners at launch. All framed as evaluation/testing rather than production rollout; the gap between “pilot” and “factory-floor economics” remains the open question.
-
Isaac Lab 3.0 Pairing: GR00T 2 ships alongside Isaac Lab 3.0, positioned as the simulated-data flywheel — NVIDIA’s end-to-end “pretrain on human video, rehearse in Isaac, deploy to the robot” pipeline.