What changed
Nvidia introduced its new AI model, GR00T 2, which transfers human movements from video footage to robotic arms without requiring direct teleoperation. The system, described as a world-action model (WAM), predicts future camera images and generates control commands for robots based on human demonstrations. Unlike its predecessor, GR00T 2 does not rely on vision-language-action translation but instead uses diffusion to predict future visual outcomes.
What this means in practice
To train a robot for a new task, users can demonstrate the steps using data gloves, hand grippers, or finger exoskeletons. The AI then generates robotic arm movements and control commands from these recordings. In lab tests, a five-fingered robot successfully performed tasks like lighting a match, slicing a banana, and tossing packages onto a conveyor belt without prior teleoperation. Success rates varied between 10 and 80 percent depending on the task. For 50 selected tasks, less than three minutes of conventional training raised success rates from 40 to 95 percent.
Limitations and availability
Nvidia paired GR00T 2 with a higher-level language model like GPT-6 Astra for complex tasks, which breaks down steps and color-codes grip points. This combination achieved 97 percent success in toy model assembly, compared to 23 percent without the language model. For bipedal robots, a slimmed-down variant, GR00T 2 Turbo, runs on the Jetson Thor onboard computer. It performed tasks like scooping popcorn and wiping liquids but succeeded in only about half of attempts for the latter. The model does not yet process direct sensor data for touch and can miss subtle errors like half-turned screws. The system also slows down due to the language model's planning time, with robots idling for a third of total time during trash sorting.
