NVIDIA GRAIL generates humanoid loco-manipulation data
Researchers present GRAIL, a digital generation pipeline that remains fully virtual until deployment by grounding video world models in 3D assets and simulator-ready scenes to synthesize humanoid loco-manipulation interactions. The system recovers metric 4D human-object interaction trajectories from privileged 3D configurations, retargets them to a humanoid robot, and trains complementary task-general trackers. The authors deploy policies trained with GRAIL-generated data on a real Unitree G1 robot for egocentric RGB-based pick-up and stair-climbing tasks. Open article →