Tech · Explainer

Nvidia GR00T 2 Carries Robot Skills Across Bodies

GR00T 2 learns from human video and transfers skills between robot bodies. Nvidia’s tests show what works, what still needs training and what developers can access.

Editorial caricature of human examples transferring to two different robot hands.
AI-generated illustration: Kunkafa

Kunkafa News · Published · 2 min read

Kareem FaridBy Discuss with the author on X

Key points

  • GR00T 2 predicts future scenes and robot movements together.
  • Nvidia reports 95% average success on 50 selected tasks after five examples per task.
  • The research page lists code for December; available N1.7 downloads are a separate release.

Nvidia’s GR00T 2 targets a costly robotics problem: teaching a skill on one body without starting again on every other machine. The company highlighted learning from people and transferring skills across hardware in its November 13 post [1].

Introduced at the Conference on Robot Learning in Austin on November 9, it belongs to Nvidia’s Cosmos-based world-action systems. The aim is to anticipate how an action changes the surroundings while generating the movement needed to produce that change [2].

How watching becomes movement

Nvidia’s paper describes a shared neural network combining camera footage, instructions and body-state data to generate future images and movements. Hardware-specific components convert the output into commands [3].

Training mixes 402,730 hours of first-person human video, 10,184 hours of handheld-gripper examples and 23,500 hours of remotely operated robot data, Nvidia says [3]. Human footage supplies experience of tasks; robot data connects that experience to executable motion.

What the robots actually did

Nvidia reports 65% success in a laboratory filter-paper task learned from humans, and 60% for YAM toy-truck assembly learned from a handheld gripper. General target-body data were still used [4].

Across 50 tasks unseen on the target robot, average success was 40% without examples, 71% with one and 95% with five. Fine-tuning took 30 minutes on one RTX 5090. Separately, 40 tasks succeeded at least once [4].

These are selected research tasks. The paper notes that the evaluation set comprises tasks learned successfully with five examples, an important boundary on the headline result [3].

From laboratory skills to industrial work

Nvidia says Foxconn is applying GR00T 2 to server manufacturing, Nexuni is testing laundry handling and Siemens is evaluating industrial uses. Its agent interface lets higher-level AI sequence skills and recover from errors. Isaac Lab 3.0 adds simulation tools for practicing manipulation before physical deployment [2].

What developers can access

The research page lists GR00T 2 code for December [4]. Nvidia separately offers a 100,000-hour Human-1M dataset subset under CC BY 4.0, including video, task descriptions and motion annotations for commercial and noncommercial use [5].

The public Isaac-GR00T repository currently documents N1.7 weights, inference and fine-tuning tools. Its code uses Apache 2.0; weights use the Nvidia Open Model License. Those downloads should not be confused with GR00T 2’s announced code release [6].

Sources: [1] Nvidia, X, 13 November 2026 · [2] Nvidia Blog, 9 November 2026 · [3] GR00T 2 paper, November 2026 · [4] Nvidia Research · [5] Human-1M dataset · [6] Isaac-GR00T repository; pages accessed 13 November 2026.

Kunkafa forecasts

NVIDIA

How far it could move, which way, by when, and how confident Kunkafa is.

Check the latest forecasts

More in Tech

From the Blog

All articles