NVIDIA Introduces GR00T 2 for Robots Learning From Human Video

NVIDIA introduced GR00T 2 on November 9, 2026, at the Conference on Robot Learning (CoRL) in Austin, and reported 40% zero-shot success on 50 tasks absent from the target robot’s training data for a model trained partly on human video.

GR00T 2 uses a 2-billion-parameter transformer built on Cosmos-Predict2.5-2B. It combines recent camera views, robot state and a written instruction to generate expected future views and the next action sequence. One pretrained checkpoint covers parallel-gripper arms, dexterous hands and a humanoid. Although NVIDIA introduces the model as open, its research page says code will be available in December.

Pretraining combines 402,730 hours of first-person human video, 10,184 hours of handheld-gripper demonstrations across 2,897 tasks and 23,500 hours of robot teleoperation. The Human-1M-100K release opens 100,000 hours of that human data, including video, language annotations and head and hand motion data, under a CC BY 4.0 license permitting commercial and noncommercial use.

Cross-embodiment tests show how skills transfer between bodies and adapt with limited robot data. NVIDIA reports 65% success for filter-paper setup learned from human demonstrations without task-specific teleoperation, and 60% for toy dump-truck assembly transferred from a handheld gripper to a YAM arm. On the Dexter evaluation, GR00T 2 completed at least one trial on 40 tasks. Average success reaches 71% with one demonstration per task and 95% with five; those five supply under three minutes of robot data per task, followed by an approximately 30-minute adapter fine-tune on one RTX 5090.

The technical whitepaper qualifies the five-demonstration result: the evaluated tasks were selected among those the model can learn with five episodes at a success rate of at least 80%, so that average is high partly by construction.

Simulation Training Turns to Precision

The simulation foundation is Isaac Lab 3.0, generally available since October 20, 2026. It provides one task API across physics, rendering and visualization backends; Newton supports rigid bodies, deformables and particles, and many workflows can run without a full Isaac Sim installation. In a vision-based dexterous-manipulation benchmark, NVIDIA reports up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX when Isaac Lab 3.0 runs with Newton.

AlphaNIST applies large-scale simulated practice to 20 NIST Assembly Task Board tasks, including parts as small as 4 millimeters. One shared policy trains with only a yes-or-no success reward. NVIDIA reports 98.2% training success for that policy, compared with 81.4% for separate per-task policies, using roughly 39 times less GPU time; these are simulation-training results.

Behind AlphaNIST is Success-Guided Sampling, developed by the University of Washington and NVIDIA. It directs practice toward configurations near the policy’s capability frontier, especially those it solves about half the time. The method’s authors report effective scaling beyond one million simulated environments and zero-shot transfer of RGB-based assembly policies to real UR5e hardware.

Benchmarks Address Industrial Variation

Isaac Lab-Arena supplies the evaluation layer. The open-source framework builds repeatable evaluations from modular components and runs them across GPU-accelerated simulated environments, with failure analysis showing which conditions need another training run. NVIDIA says its 0.4 developer preview is arriving at CoRL.

DexBench v1.0, developed by RLWRLD with NVIDIA, contains 15 tasks across retail, logistics, assembly-line and hotel food-and-beverage scenes. VABAR, from UC Berkeley and NVIDIA, covers tool sorting and hanging, gear meshing, syringe packing, and cable insertion and routing, supported by a 33-skill library. VABAR’s project page says its two current difficulty levels form the validation set and a harder test set is still coming.

Agents Add Recovery and New Constraints

UC Berkeley’s Graph-as-Policy framework uses multiple agents to compose an interpretable directed graph of robot skills, rehearse it in simulation and revise its structure and parameters after failures. The project reports that ten rehearsal iterations lifted its Make Popcorn result from 33% initially to 94% in simulation and 90%, or 18 of 20 runs, on a real robot. Its current benchmarks emphasize quasi-static manipulation; deformables, dynamics and moving targets remain future work.

GR00T 2 uses a reasoning agent to sequence skills, monitor progress and invoke recovery routines when parts drop or work stalls, while the model handles contact-rich manipulation. With corrective data and that agent, NVIDIA reports 98 successful toy-plane assemblies, 97 toy-tractor assemblies and 97 Hot Wheels assemblies in 100 trials per toy, each starting from randomly dumped parts. Supervised fine-tuning alone produced 69, 51 and 23, respectively.

The whitepaper separately reports roughly 70% average success on its long-horizon post-training benchmarks. In visually steered trash sorting, the robot waited for the agent during a median 38% of an episode, exposing the latency cost of agent oversight.

NVIDIA says Nexuni is testing GR00T 2 for laundry handling, Foxconn is applying it to server-manufacturing automation and Siemens is evaluating industrial use cases. The disclosed work is testing and evaluation; no production rollout is announced.

The whitepaper’s conclusion divides reliability among pretraining for breadth, corrective demonstrations and reinforcement learning for precision, and agents for memory, verification and recovery. For industrial users, the useful measure is therefore the full workflow around GR00T 2 and the supervision each task consumes. The open deployment challenge is to preserve the reported reliability while reducing the per-task adaptation and agent waiting time documented in NVIDIA’s own results.