Therapists' Force-Profile Teach-and-Mimic Approach for Upper-Limb Rehabilitation Exoskeletons

In this work, we propose a framework enabling upper-limb rehabilitation exoskeletons to mimic the personalised haptic guidance of therapists. Current exoskeletons face acceptability issues as they limit physical interaction between clinicians and patients and offer only predefined levels of support that cannot be tuned during the movements, when needed. To increase acceptance, we first developed a method to estimate the therapist's force contribution while manipulating a patient's arm using an upper-limb exoskeleton. We achieved a precision of $0.31Nm$ without using direct sensors. Then, we exploited the Learning-by-demonstration paradigm to learn from the therapist's interactions. Single-joint experiments on ANYexo demonstrate that our framework, applying the Vector-search approach, can record the joint-level therapist's interaction forces during simple tasks, link them to the kinematics of the robot, and then provide support to the user's limb. The support is coherent with what is learnt and changes with the real-time arm kinematic configuration of the robot, assisting whatever movement the patient executes in the end-effector space without the need for manual regulation. In this way, robotic therapy sessions can exploit therapists' expertise while reducing their manual workload.