A Framework of Rehabilitation-assisted Robot Skill Representation, Learning, and Modulation via Manifold-Mappings and Gaussian Processes

摘要

Stroke survivors usually have dyskinesia, who have an urgent need for rehabilitation-assist training. To reduce the labor of rehabilitation therapists, this paper attempts to investigate an effective rehabilitation-assisted robot skill acquisition framework, which is inspired by the scheme of robot learning from demonstration (LfD). Since most of the current LfD methods were implemented with rigorous assumptions that the considering motion features are only represented on an individual manifold. Meanwhile, despite many advancements that have been achieved on time-position trajectories and position-velocity trajectories, those methods are restricted to Euclidean space and can not be applied to learn those dexterous and compliant rehabilitation-assisted robot skills such as position-orientation trajectories and force-stiffness trajectories, etc. In this paper, we propose a novel skill acquisition framework for rehabilitation-assisted robot using manifold-mappings and Gaussian processes, which allows the robot to 1) simultaneously considering the robot position, orientation, force as well as stiffness by manifold-mappings among d-dimensional Euclidean space <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{R}^{d}$</tex> , special orthogonal group <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$S\mathcal{O}$</tex> (3), and Riemannian space <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{M}$</tex> , respectively, which resulting in accurate motion and compliant behavior; 2) retrieving skill representation by encap-sulating the variability of multiple high-dimensional demon-strations that with input-dependent noises; 3) implementing the via-points-based trajectory modulation by considering task constraints or environmental changes. To simplify the writing, we named the proposed framework as Multi-motion Features Fusion-based Robot Skill Learning (MF <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> RoSL). To effectively evaluate the effectiveness of our proposed method, an upper limb rehabilitation training system with a collaborative Kinova robot is developed. The training exercises of our system are determined according to the Brunnstrom therapeutic approach to the management of hemiplegic patients, including the 3-DoFs movement of the shoulder joint and a 7-DoF movement of an insertion/extraction task for assessing the activities of daily living (ADL). Results indicate that our proposed MF <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> RoSL method allows the robot to learn rehabilitation skills from the therapist and can be rapidly adapted to new patients.