DEV Community

News
News

Posted on

NVIDIA GR00T 2 learns robot skills from 400000 hours of human video

Infographic cover

๐Ÿ“… November 10, 2026 โ€ข NVIDIA Research

NVIDIA published a research report on GR00T 2, an AI foundation model built to cut teleoperation in robot training, per NVIDIA via The Decoder on November 10, 2026. GR00T 2 is a "world-action model" โ€” no longer a vision-language-action model like GR00T 1. It simultaneously computes the next movement and predicts a short video of the immediate future using diffusion. The model was trained on a cluster with more than 1000 GB300 chips and about 400000 hours of head-mounted camera video of people performing tasks.

A person demonstrates the steps using data gloves, specialized hand grippers or a finger exoskeleton, and combined with general motion data of the target robot the model generates the robotic arm inside its own video prediction and derives control commands, per The Decoder. In lab tests, a robot with five-fingered hands lit a match, sliced a banana and tossed packages onto a conveyor belt without being teleoperated for those specific tasks โ€” success rates varied between 10 and 80 percent depending on the task.

Nail hammering was learned from a single video of a human hand via simulation, per The Decoder. On 50 selected tasks where the model performed well after a few examples, less than 3 minutes of conventional training raised success from 40 to 95 percent. For longer assembly work, NVIDIA pairs GR00T 2 with a higher-level language model like OpenAI GPT-6 Astra: it breaks the task into steps, color-codes grip points in the camera image and launches helper routines. With that stack, toy assembly succeeded in at least 97 percent of trials, versus 23 out of 100 attempts without it.

A slimmed-down variant, GR00T 2 Turbo, targets bipedal robots and runs directly on the Jetson Thor onboard computer. In tests the humanoid scooped popcorn, cleared tables and wiped liquids โ€” the last task succeeded in just under half of attempts. Stated hurdles remain: the model does not yet process touch sensor data and misses subtle errors like half-turned screws, and the planning model slows the system โ€” during trash sorting the robot idled about one third of the time, waiting over 7 seconds on average for the next plan, per The Decoder.

Key figures

IMPACT: teleoperation time drops from hours of manual control to short demonstrations, and skills transfer between robot body types. Human video becomes scalable training data for manipulation. Reliability still depends on the slower planning layer โ€” during trash sorting the robot idled about one third of the time โ€” so automation engineers should plan for idle windows and check the success rate of each task before production use.

IMPACT checklist


Sources:


Continue the conversation: https://t.me/AIAutomationArchitecture ยท Bluesky

AI Automation

Top comments (0)