NVIDIA launched GR00T 2, a world action model built on NVIDIA Cosmos that predicts future video and robot actions together, at the Conference on Robot Learning (CoRL 2026) in Austin