NVIDIA's GR00T 2 Learns Robot Skills From 400,000 Hours of 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. The number matters because general-purpose robot learning has historically required task-specific demonstrations on the exact hardware; a model that transfers from human first-person footage to unseen robot tasks attacks that cost directly.

GR00T 2 is 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. NVIDIA presents the model as open, though its research page states code will be available in December.

How much data does GR00T 2 need?

Pretraining draws on three large corpora: 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. Alongside the model, NVIDIA is releasing Human-1M-100K, a dataset of 100,000 hours from that human video collection, including video, language annotations and head and hand motion data, under a CC BY 4.0 license that permits commercial and noncommercial use.

Cross-embodiment tests (evaluations where a skill learned on one robot body transfers to another) show how the model adapts with limited robot data. NVIDIA reports:

• 65% success for filter-paper setup learned from human demonstrations without task-specific teleoperation
• 60% for toy dump-truck assembly transferred from a handheld gripper to a YAM arm
• On the Dexter evaluation, completion of at least one trial on 40 tasks

Average success on Dexter reaches 71% with one demonstration per task and 95% with five. Those five demonstrations supply under three minutes of robot data per task, followed by an approximately 30-minute adapter fine-tune on a single RTX 5090.

The technical whitepaper qualifies the headline five-demonstration figure. The evaluated tasks were selected among those the model can learn with five episodes at a success rate of at least 80%, so that 95% average is high partly by construction. Readers comparing robot-learning claims across labs should note the selection effect.

What changes in simulation training?

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, its physics engine, 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 shared policy, compared with 81.4% for separate per-task policies, using roughly 39 times less GPU time. These are simulation-training results, not real-world success rates.

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

Which benchmarks cover 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.

Two additional benchmark suites target industrial settings. 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.

How do agents improve reliability?

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 pairs a reasoning agent with the manipulation model. The agent sequences skills, monitors progress and invokes 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. One measurement exposes the cost of agent oversight: in visually steered trash sorting, the robot waited for the agent during a median 38% of an episode.

Who is testing it?

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

What is the open question?

The whitepaper's conclusion divides reliability among three components: 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 — not the zero-shot or five-shot numbers alone.

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. The median 38% agent-wait share in trash sorting and the 30-minute per-task fine-tune give prospective users concrete thresholds to beat.