Nvidia's model GR00T 2 watches humans on video, then imagines itself as the robot

Nvidia has introduced GR00T 2, an AI model that translates human hand movements observed in video footage into instructions for robotic arms. By eliminating the need for manual teleoperation, the system aims to streamline how robots learn new tasks. While the model has been trained on 400,000 hours of video, its current performance is inconsistent, with success rates fluctuating between 10 and 80 percent. This technology represents a shift toward scalable robot training, though its reliability remains a significant barrier to practical implementation.