Nvidia's GR00T 2 model learns tasks by watching humans, predicts robotic movements

nvidia's gr00t 2 is a new ai foundation model designed to shorten robotic training processes. it works as a world-action model, simultaneously calculating the next movement and a short video of the immediate future. gr00t 2 was trained on a cluster with over 1,000 gb300 chips and 400,000 hours of video footage. in lab tests, a robot with five-fingered hands performed tasks like lighting a match or slicing a banana without prior teleoperation. success rates varied between 10 and 80 percent depending on the task.