KinoGen turns generated video into motion references for a humanoid robot

Researchers report deploying control policies trained on generated motion references on a Unitree G1 humanoid with dexterous hands. Their KinoGen preprint describes a method for creating robot movements that combine locomotion with object handling, without first translating movements from a human body model. The system conditions a video generator on simulated views of the robot and object, then fits those models to generated video from two viewpoints. Across eight tasks, the references preserved over 90% of annotated contacts and showed less object penetration and foot skating than video-based baselines. The comparative results concern motion references.