上肢康复机器人主动康复策略的学习及迁移方法
Learning and Transfer Methods for Active Rehabilitation Strategy of Upper-limb Rehabilitation Robot
-
摘要: 为实现脑卒中患者上肢的主动康复, 提出了基于强化学习的反馈控制方法, 将训练过的控制策略迁移至实体环境。首先通过分析人体上肢运动, 建立了人体上肢等效模型, 并提出了在关节空间拟合上肢随意运动的随机轨迹规划方法; 其次, 基于等效模型建立了人机耦合系统的仿真环境, 通过人体及机器人运动学分析将人机耦合系统的完整状态映射为可观测状态, 以实现控制策略训练; 最后, 为了将仿真中训练的控制策略迁移至机器人样机, 利用反馈控制策略对齐环境间的特征分布, 避免了策略在不同环境间的负迁移。仿真结果表明, 基于强化学习的控制策略可实现脑卒中患者的主动康复; 将仿真中训练的控制策略迁移至现实机器人, 实验结果验证了康复策略的学习与迁移方法的有效性, 为运动障碍患者的主动康复提供了一种新的技术解决方案。Abstract: To achieve active upper-limb rehabilitation for stroke patients, a feedback control strategy based on reinforcement learning is proposed, to transfer the control strategy to the physical environment after training. Firstly, human upper-limb movement is analyzed to establish an equivalent model, and a random trajectory planning method is proposed to mimic voluntary movement in joint space. Subsequently, a simulation environment for man-robot system is established based on the equivalent model. To train the control strategy, the complete states of the system are mapped to observable states through kinematic analysis on the human body and robot. Finally, a feedback control strategy is used to align the feature distribution between environments, in order to transfer the trained control strategy to the robot and avoid negative transfer. Simulation results demonstrate the effectiveness of the reinforcement learning-based control strategy in accomplishing active rehabilitation tasks for stroke patients. The control strategy trained in simulation is transferred to the real robot, and the experimental results verify the effectiveness of the learning and transfer method of rehabilitation strategy, which provides a new technical solution for the active rehabilitation of movement disorder patients.