“Sticky Hands”: Learning and Generalization for Cooperative Physical Interactions With a Humanoid Robot

"Sticky Hands" is a physical game for two people involving gentle contact with the hands. The aim is to develop relaxed and elegant motion together, achieve physical sensitivity-improving reactions, and experience an interaction at an intimate yet comfortable level for spiritual development and physical relaxation. We developed a control system for a humanoid robot allowing it to play Sticky Hands with a human partner. We present a real implementation including a physical system, robot control, and a motion learning algorithm based on a generalizable intelligent system capable itself of generalizing observed trajectories' translation, orientation, scale and velocity to new data, operating with scalable speed and storage efficiency bounds, and coping with contact trajectories that evolve over time. Our robot control is capable of physical cooperation in a force domain, using minimal sensor input. We analyze robot-human interaction and relate characteristics of our motion learning algorithm with recorded motion profiles. We discuss our results in the context of realistic motion generation and present a theoretical discussion of stylistic and affective motion generation based on, and motivating cross-disciplinary research in computer graphics, human motion production and motion perception.