NVIDIA's GR00T 2 hits 95% on robot tasks after five demos

NVIDIA's new robot-control model GR00T 2 averaged 95% success across 50 manipulation tasks after training on just five short demonstrations per task, the company reported Nov. 9 at the Conference on Robot Learning in Austin, Texas. Before those demonstrations, the same model averaged 40% on tasks it had never practiced. Both figures come from NVIDIA's own evaluation, and the company concedes the headline number is partly an artifact of test design.

The caveat sits in the white paper itself. NVIDIA selected only tasks on which the model could reach at least 80% success after five demonstrations, a choice that makes the five-demonstration result "high partly by construction," the document states. Before task-specific training, the model completed 40 of the 50 tasks at least once. After training, it succeeded on every trial of 37 tasks. For R&D managers weighing foundation models for automation pilots, the relevant question is not the 95% but which tasks cleared the 80% selection bar and why.

How the model works

NVIDIA calls GR00T 2 a "world action model" because it jointly predicts future video and future robot movements. It builds on Cosmos-Predict2.5-2B, a video-generation model in NVIDIA's Cosmos family for physical AI (NVIDIA's term for AI that acts in the physical world). The system takes recent camera footage, a language instruction and the robot's current configuration, then generates video of how the scene could change alongside a sequence of movements. The robot executes those movements, absorbs new observations and repeats the cycle. A smaller, faster variant, GR00T 2 Turbo, also controlled a humanoid robot serving popcorn in NVIDIA's demonstrations.

The rollout accompanies Isaac Lab 3.0, the latest version of NVIDIA's open-source robot learning framework, as The Robot Report noted.

The chemistry-lab stress test

A separate experiment addressed a question lab directors will recognize: can robots learn bench procedures from human demonstration? Researchers gave Dexter, a two-armed test robot fitted with five-fingered Sharpa hands, 100 human demonstrations for each of three chemistry-lab tasks, recorded via head and wrist cameras, motion-tracking gloves and wrist trackers. They combined these with general practice data gathered through teleoperation (remotely controlling the robot), but supplied no teleoperated demonstrations of the target procedures.

Results before any robot-specific demonstrations:

• Filter-paper placement in a funnel, including unfolding, wetting and bottle return: 13 of 20 trials, or 65%
• Dispensing liquid into a beaker with a dropper: 30%
• Dipping a swab in water and dabbing a microscope slide: 10%

Adding 10 teleoperated demonstrations per task lifted success to 95% for filter-paper setup, 70% for dispensing and 50% for slide dabbing. The comparison is instructive. A version pretrained only on data from the target robot reached just 45%, 40% and 30% after identical task training. GR00T 1.7, NVIDIA's previous model, completed none of the chemistry-lab trials under any condition. All models received the same demonstrations, practice data and training procedure; each result reflects 20 trials, and any human intervention counted as failure.

Longer tasks and the missing rivals

On seven longer, multistep tasks using dexterous hands — including filtering crude salt and replacing a computer memory module — GR00T 2 averaged 68.6% success. GR00T 1.7 averaged 0.7%. π0.5, a 2025 model from robotics AI startup Physical Intelligence, scored 0%. Across six tasks using a robot arm with a two-finger gripper, GR00T 2 averaged 71.7% and GR00T 2 Turbo 65%; the five baseline models peaked at 12.5%, per Table 1 of the white paper.

Two qualifications temper those gaps. The reported truck-assembly result used easier starting arrangements for GR00T 2 and GR00T 2 Turbo; on the original arrangements, GR00T 2 completed seven of 20 trials (35%) versus the reported 65%. And the comparison set omits the current leaders: Physical Intelligence released π0.7 on April 16, and Google DeepMind announced Gemini Robotics 2 on July 30. Neither appears in these benchmarks.

The open question is how GR00T 2's five-demonstration efficiency holds on tasks NVIDIA did not pre-screen, and against models released this year.