Duke University Researchers Enable Humanoid Robots to Dance with Humans Using CoDance Framework

Researchers at Duke University developed CoDance, a framework allowing humanoid robots to dance with human partners while maintaining their own style and responding compliantly to physical contact. By processing a short video clip and applying multi-link compliance enhancement, the system generates training data that includes inferred force information without requiring additional sensors. Trained on sparse keypoints of the partner's body, the robot successfully completed multiple dance sequences lasting up to 38.9 seconds on a Unitree G1 robot in real-world tests without falling, though it remains dependent on reference actions and cannot yet handle complex interactions like rotations or lifts.