NVIDIA's GR00T 2 Teaches Robots New Skills by Watching Humans

Robots that can learn new tasks from just a handful of human demonstrations moved closer to reality on November 9, 2026, when NVIDIA presented GR00T 2 at the Conference on Robot Learning (CoRL) in Austin, Texas.

The model is a 2-billion-parameter transformer built on NVIDIA's Cosmos world-simulation architecture. Give it camera views, the robot's joint states, and a plain-text instruction, and it predicts what the world should look like next — then acts to make that happen. One checkpoint covers parallel-gripper arms, dexterous hands, and full humanoid robots simultaneously, so a lab does not need separate models for each hardware platform.

Learning From Five Videos

The headline result, described in NVIDIA's CoRL announcement, is that GR00T 2 reaches a 95% success rate across a 50-task benchmark after just five human demonstrations per task. With a single demonstration it still succeeds about 71% of the time, and with zero demonstrations it completes 40 of the 50 tasks at least once — tasks the target robot never trained on. That zero-shot generalization is the number to watch, because it hints at a model that can be dropped into new environments without an expensive data-collection phase.

The training pipeline runs inside Isaac Lab 3.0, NVIDIA's GPU-accelerated robot simulator, which shipped alongside GR00T 2. Two new open benchmarks — DexBench for dexterous manipulation and VABAR for industrial tasks — provide standardized tests so different research groups can compare results on equal footing.

What "Open" Means Here

NVIDIA is calling GR00T 2 open, but as of the CoRL announcement no weights, model card, or license have been published. The AI Insider's coverage notes that NVIDIA's research page promises code availability in December. That is an important caveat: open in intention is not the same as open in practice, and the robotics community will be watching the December release closely to see what access is actually granted.

Why It Matters

The race to build a "foundation model for robots" — a single pretrained brain that any robotic body can fine-tune cheaply — has been underway since large language models proved that the approach works for text. GR00T 2 is NVIDIA's most concrete public step in that direction. The combination of a world-model backbone (predicting future states, not just actions), a unified checkpoint for different hardware, and strong few-shot performance makes it one of the more credible entries in that race so far.

For robotics researchers, the December code drop will be the real test. If the weights are genuinely usable and the license is permissive, GR00T 2 could become the base checkpoint that labs build from rather than train from scratch — compressing years of data collection into an afternoon of fine-tuning.

NVIDIA also announced that GR00T 2 results were presented in a paper at CoRL 2026, though the paper was not publicly linked at the time of announcement. The Unite.AI writeup and Blockchain News both have solid summaries of the technical architecture if you want to dig further.