NVIDIA’s GR00T 2 reaches 40% zero-shot success on 50 unseen tasks

By Daniel Okafor

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

What does GR00T 2 learn from human video?

NVIDIA’s whitepaper specifies 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 checkpoint covers parallel-gripper arms, dexterous hands and a humanoid.

NVIDIA’s pretraining mix:

• 402,730 hours of first-person human video.
• 10,184 hours of handheld-gripper demonstrations across 2,897 tasks.
• 23,500 hours of robot teleoperation.

NVIDIA’s Human-1M-100K release opens 100,000 hours of human data. It includes video, language annotations, head and hand motion data. Its CC BY 4.0 license permits commercial and noncommercial use.

NVIDIA called the model “open,” but its research page says code will be available in December.

How well do skills transfer across robots?

NVIDIA reported 65% success on filter-paper setup learned from human demonstrations. That setup required no task-specific teleoperation. The company also reported 60% success on toy dump-truck assembly. That skill transferred from a handheld gripper to a YAM arm.

On the Dexter evaluation, NVIDIA reports that GR00T 2 completed at least one trial on 40 tasks. Average success reached 71% with one demonstration per task. It reached 95% with five. Those five demonstrations supplied under three minutes of robot data per task. That step used an approximately 30-minute adapter fine-tune on one RTX 5090.

The technical whitepaper qualifies the result. Evaluators selected tasks the model could learn in five episodes at an 80% success rate or better. The average is “high partly by construction,” the whitepaper says, because of that selection.

What does Isaac Lab 3.0 change?

NVIDIA says Isaac Lab 3.0 has been generally available since October 20, 2026. The platform provides one task API across physics, rendering and visualization backends. Newton supports rigid bodies, deformables and particles. Many workflows can run without a full Isaac Sim installation.

In a vision-based dexterous-manipulation benchmark, NVIDIA reported up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX when Isaac Lab 3.0 runs with Newton.

Can one shared policy replace task-specific training?

NVIDIA’s AlphaNIST project applies large-scale simulated practice to 20 NIST Assembly Task Board tasks, including parts as small as 4 millimeters. It trains one shared policy with only a yes-or-no success reward.

NVIDIA reports 98.2% training success for that policy, against 81.4% for separate per-task policies. The shared approach used roughly 39 times less GPU time. These are simulation-training results.

Behind AlphaNIST, the University of Washington and NVIDIA developed Success-Guided Sampling. It directs practice toward configurations near the policy’s capability frontier, especially those solved 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.

What do the new benchmarks measure?

NVIDIA’s Isaac Lab-Arena supplies the evaluation layer. The company describes the open-source framework as building repeatable evaluations from modular components across GPU-accelerated simulated environments. Its failure analysis identifies conditions that need another training run. NVIDIA says version 0.4 will arrive as a developer preview 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. It includes a 33-skill library.

VABAR’s project page says two current difficulty levels form the validation set. A harder test set remains forthcoming.

Can agents recover from failed steps?

UC Berkeley’s Graph-as-Policy framework composes an interpretable directed graph of robot skills with multiple agents. It rehearses the graph in simulation. After failures, it revises the graph’s structure and parameters.

The UC Berkeley project reports that ten rehearsal iterations lifted Make Popcorn from 33% to 94% in simulation. On a real robot, it reached 90%, or 18 of 20 runs. Current benchmarks emphasize quasi-static manipulation. Deformables, dynamics and moving targets remain future work.

GR00T 2 adds a reasoning agent that sequences skills and monitors progress. It invokes recovery routines when parts drop or work stalls. The underlying model handles contact-rich manipulation. With corrective data and the agent, NVIDIA reports 98 successful toy-plane assemblies, 97 toy-tractor assemblies and 97 Hot Wheels assemblies. Each covers 100 trials per toy, starting from randomly dumped parts. Supervised fine-tuning alone produced 69, 51 and 23 successes, respectively.

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

What remains before industrial deployment?

According to NVIDIA, Nexuni is testing GR00T 2 for laundry handling. Foxconn is applying it to server-manufacturing automation. Siemens is evaluating industrial use cases. All disclosed work remains testing and evaluation. No production rollout has been announced.

The whitepaper divides reliability among pretraining for breadth, corrective demonstrations and reinforcement learning for precision, and agents for memory, verification and recovery. For industrial users, the practical measure is the full workflow around GR00T 2 and the supervision each task consumes. NVIDIA’s results frame the next test: preserve reported reliability while reducing per-task adaptation and agent waiting time.