Nvidia's GR00T 2 turns human video into robot actions
Original title: Nvidia's model GR00T 2 watches humans on video, then imagines itself as the robot
Lab success rates ranged from 10 to 80 percent; with a planner, toy assembly reached at least 97 percent of trials.
Nvidia's research report introduces GR00T 2, a world-action model that predicts the next robot movement and a short future camera video instead of mapping images and language directly to motors like GR00T 1. Training used a cluster of more than a thousand GB300 chips and about 400,000 hours of head-mounted human task video. New tasks can be demonstrated with data gloves, grippers, or a finger exoskeleton. The team says the model still lacks touch sensing, can miss subtle errors, and can idle while a higher-level language model plans.
Unconfirmed: Success rates and training claims come from Nvidia's lab report and are not independently verified.
- For 50 tasks, under three minutes of training raised success from 40 to 95 percent.
- Without the language model, toy-car assembly succeeded in 23 of 100 attempts.
- GR00T 2 Turbo on Jetson Thor wiped liquids in just under half of attempts.
- During trash sorting the robot waited more than seven seconds on average for planning.
- Vendors
- Nvidia
- Products
- Jetson Thor
- Organizations
- OpenAI
- AI models
- GR00T 2GR00T 2 TurboGPT-6 Astra
From the original article
3 min read at the-decoder.com
Nvidia's new AI model GR00T 2 is designed to teach robots new tasks without laborious teleoperation. Instead, the system transfers human hand movements from videos to robotic arms, trained on about 400,000 hours of video footage. Success rates still vary widely, though, ranging from 10 to 80 percent depending on the task.
By Maximilian Schreiner · the-decoder.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.