At the 7th International Workshop on Active Inference (IWAI 2026), held October 14–16, 2026 at the Spanish National Research Council (CSIC) in Madrid, I presented “Toward Human-Robot Synergy Beyond Individual Capability: Personalized Assistance for Motor Learning as a Foundation for Adaptive Collaboration” as both an oral talk and a poster. This is joint work with Shunichi Kasahara at Sony CSL.
Our research question is how the balance between a robot's prior-driven behavior and its sensory adaptation influences the transfer of motor skills to humans through physical human-robot interaction. Haptic interaction can convey intention, yet too much guidance is known to hinder learning, so the key is how to design the space in between.
As our testbed, we built a cooperative seesaw task in which a human and a robot hold opposite ends of a board and tilt it to control a ball. The robot is driven by PV-RNN, a predictive-coding-inspired variational RNN, which updates its internal state and motion predictions online as human intervention changes what it observes.
Our focus is PV-RNN's meta-prior w. A large w makes the robot preserve its own predictions, closer to a leader-style behavior, while a small w makes it adapt its internal state to the partner, closer to a follower-style behavior. Next, we will vary w as an experimental condition and measure its effects on task performance, sense of agency, and skill retention and transfer, with the eventual goal of adapting w itself to each person's strengths and to the situation.