In early November 2026, at the Conference on Robot Learning in Austin, NVIDIA introduced GR00T 2, a new generation of robot foundation models designed to make skill acquisition far more efficient.
On 50 tasks the target robot had never practiced (seen only in human video or data from other robots), the model reached roughly 40% average success zero-shot. One demonstration per task lifts that to about 71%, and five demonstrations under three minutes of robot data raise it to 95% after a short fine-tune.
This is not incremental progress on vision-language-action (VLA) models. GR00T 2 is a world-action model (WAM). It predicts future video frames and the robot’s next actions together in a single diffusion process, learning how physical interactions unfold rather than mapping observations and language directly to motor commands.
How the Model Works
GR00T 2 is built on NVIDIA’s Cosmos platform. A diffusion transformer with roughly 2 billion parameters. GROOT 2 takes recent multi-camera video typically the last 0.8 seconds from head and wrist views, the robot’s current state, and a language instruction. It jointly denoises a short future video sequence and an action chunk. Small per-robot heads then convert the planned actions into the specific joint commands needed for parallel grippers, multi-fingered hands, or a full humanoid body.
Because the model imagines the visual outcome of its actions, it can extract useful physics and manipulation knowledge from pure human video. This is the foundation of its cross-embodiment transfer capability, which NVIDIA calls X-Adapt.
Training Data Made up of Mostly Human Video
The pretraining mixture is dominated by human data:
· 402,730 hours of egocentric video of people working culminating in roughly 92% of the total mix.
· 10,184 hours of handheld UMI gripper demonstrations across nearly 2,900 tasks.
· 23,500 hours of robot teleoperation spanning arms (YAM, Franka), dexterous hands (Sharpa, Wuji), and the Unitree G1 humanoid.
NVIDIA released 100,000 hours of the human egocentric video openly as Human-1M-100K on Hugging Face under a permissive license. Scaling experiments showed that increasing the share of human video in pretraining steadily improved real-robot task progress on downstream benchmarks.
Few-shot Results and Transfer examples
The 50-task evaluation used a bimanual robot with two high-DoF Sharpa hands. Tasks were deliberately chosen from those appearing in human video or other robots’ data but never demonstrated on the target platform. Adaptation uses a lightweight LoRA fine-tune that takes about 30 minutes on a single RTX 5090.
Concrete transfer examples include:
· A human demonstration of lining a laboratory funnel with filter paper and wetting it; a five-fingered robot then succeeds in a majority of trials with no teleoperated data of its own for that task.
· Handheld UMI gripper demonstrations of assembling a toy dump truck; a different robot arm (YAM) completes the multi-minute sequence end-to-end in a substantial fraction of trials.
· Wearable DexUMI hand data enabling long-horizon bimanual skills such as preparing breakfast or lighting a lamp with a match.
With modest additional post-training, the model handles longer, more dexterous sequences including memory module insertion, package labeling, toy assembly that earlier models struggled with. Real-robot reinforcement learning further boosts contact-rich insertion tasks into the high 80s to mid-90s percent success range.
A lighter variant, GR00T 2 Turbo, runs onboard a Unitree G1 humanoid’s Jetson Thor at 124 ms per prediction and supports whole-body control for tasks such as serving popcorn or clearing a table.
Broader Context and Ecosystem
GR00T 2 was announced alongside Isaac Lab 3.0, which improves simulation throughput up to 2.5× on certain dexterity benchmarks, supports interchangeable physics solvers (including Newton for better soft-body and multi-joint handling), and adds tooling for agent-assisted environment setup and reinforcement post-training of foundation models.
NVIDIA also highlighted new industrial benchmarks and partner activity such as Foxconn exploring server manufacturing automation, Siemens evaluating industrial uses, others testing laundry handling and related applications.
Code and model weights are scheduled for December 2026. The company positions the release as open, though final licensing details will matter for commercial adoption.
Why it Matters
Robot skill learning has long been bottlenecked by the cost and time of collecting high-quality teleoperated demonstrations on the exact target hardware. By shifting the bulk of learning onto abundant human video and enabling rapid few-shot adaptation, GR00T 2 points toward a more scalable path. The ability for a higher-level reasoning agent to call the model via API, sequence skills, and recover from errors further reduces the need for fully scripted behaviors.
Important caveats remain. The 95% figure is NVIDIA’s own evaluation on a specific set of 50 tasks; independent verification and results on customer-chosen industrial tasks will be the real test. Success rates on longer-horizon or highly variable real-world conditions will likely be lower than the controlled numbers. Embodiment gaps, safety, and reliability in unstructured environments still require careful engineering.
Nevertheless, the architectural shift to joint video-and-action prediction, the heavy use of human data, and the demonstrated few-shot transfer represent a meaningful step toward generalist robot policies that can acquire new skills the way people do—by watching and then practicing a handful of times. For robotics researchers, integrators, and companies building next-generation automation, GR00T 2 is one of the more consequential releases of 2026.
