GR00T 2

GR00T 2 is NVIDIA’s open frontier world action model, introduced at CoRL 2026 and built on the NVIDIA Cosmos stack for humanoid robots and dexterous manipulation.

It is trained on a broad mix of robot and human data—including roughly 100,000 hours of egocentric footage—so it can complete many manipulation tasks zero-shot and transfer skills across different robot bodies. NVIDIA positions it among the leading vision-language-action (VLA) / world-action models on RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces. Early adopters cited in launch coverage include Nexuni, Foxconn, Siemens, and MicroAGI.

Highlights:

• Announced at CoRL 2026; NVIDIA robotics simulation blog: https://blogs.nvidia.com/blog/robotics-simulation-advancements-corl/ ;
• Cosmos-based open world action / robot foundation model aimed at dexterous manipulation;
• Skill transfer across morphologies plus an API for agentic skill sequencing;
• Strong benchmark standing on RoboCasa / RoboLab-family suites;
• Already referenced in industrial and research pilots (Nexuni, Foxconn, Siemens, MicroAGI).

Why it matters

Humanoid and multi-robot fleets need a shared “world action” prior that can plan and execute contact-rich skills without retraining from scratch for every body. GR00T 2 targets that gap: a Cosmos-backed open model that absorbs egocentric human and robot data, then generalizes zero-shot and across embodiments. That matters for labs and manufacturers who want one foundation model to sequence skills through an agent API rather than hand-tuning separate policies per robot.

Public launch materials emphasize top-tier placement on RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces—suites that stress household and lab-style manipulation under visual variation. The combination of open weights (as framed by NVIDIA), skill-transfer claims, and named industrial users is why the radar scored this term highly in the CoRL window.

Limits

Benchmark leadership on selected suites does not guarantee every real-world contact task. Embodiment transfer, safety envelopes, and sim-to-real gaps still depend on the integrator’s stack (Isaac / Cosmos tooling, controllers, and sensors). Confirm license terms, supported robot APIs, and current checkpoint availability on NVIDIA’s channels before production. Do not treat unpublished parameter counts or private eval scores as facts.

1. Read NVIDIA’s CoRL / robotics simulation post at https://blogs.nvidia.com/blog/robotics-simulation-advancements-corl/ and follow links to the current GR00T 2 model cards and Cosmos docs;
2. Evaluate on RoboCasa / Isaac Lab–style setups if you already use NVIDIA’s robotics stack;
3. Use the agentic skill-sequencing API for multi-step manipulation rather than single-shot demos;
4. Before fleet rollout, verify license, supported embodiments, and safety monitors; do not assume unpublished quotas or private benchmark numbers.