Contextual Policy Search for Task-Level Adaptation in Physical Human-Robot Interaction

Physical human-robot interaction controllers play a crucial role in various robotic applications, enabling robots to act as compliant and intelligent collaborators or assistants. While significant progress has been made in force-level and motion-level adaptations, task-level adaptation in unseen scenarios remains underdeveloped. To address this gap, we propose a learning framework with three core contributions to enable robots to modulate interactive behaviours for a family of tasks defined by context variables. First, a lower-level interactive policy is developed based on impedance regulation and a safety-stop mode, allowing the robot to safely and compliantly interact with humans by interpreting their motion preferences and motion intentions. Second, a linear Gaussian contextual policy is formulated as the higher level policy to learn the mapping from the context space to the parameter space of the lower level interactive policy. Third, a latent interactive space is constructed based on the detected human motion intentions, enhancing sample efficiency in task-level adaptation learning. Without loss of generalization, this study focuses on the applications of robot-aided training and rehabilitation. The effectiveness of the proposed learning framework is demonstrated by a human subject study using a wrist robot for point-reaching tasks. Furthermore, two commonly used contextual policy learning methods have validated the proposed framework's ability to improve sample efficiency.