CoDance: Learning Reactive and Compliant Human-Humanoid Interaction from Video
CoDance turns a single video of partnered dancers into training data for a humanoid that can keep two-hand contact while adapting to a human partner.
The framework retargets one dancer as the robot reference and the other as the moving partner. It adds force-aware compliance data by adjusting the robot motion under structured forces at both hands. In simulation, the learned policies adapted footsteps to partner changes and reproduced about 80% of the augmented wrist displacement. On hardware, the humanoid sustained two-hand dancing with a person, including repeated forward and backward transitions. HF Daily Papers' note
The framework retargets one dancer as the robot reference and the other as the moving partner. It adds force-aware compliance data by adjusting the robot motion under structured forces at both hands. In simulation, the learned policies adapted footsteps to partner changes and reproduced about 80% of the augmented wrist displacement. On hardware, the humanoid sustained two-hand dancing with a person, including repeated forward and backward transitions. HF Daily Papers' note
score 5