Why it matters
Robot policies need experience grounded in physical actions. Useful datasets can include demonstrations, robot trajectories, synthetic experience and human videos. NVIDIA describes GR00T as trained on real, synthetic and internet-scale video data, illustrating how multimodal data sources are being combined.