| 摘 要: 针对下肢外骨骼的传统控制策略难以动态适配不同康复阶段，且存在人机交互安全风险。提出一种带物理约束的改进SAC强化学习自适应导纳控制算法，以实现外骨骼轨迹跟踪与人机交互柔顺性的协同优化。引入约束马尔可夫决策过程，结合拉格朗日乘子法设置物理硬约束。结果表明，该算法在1000个训练Episode内稳定收敛，约束违反代价最终趋近于0。在被动训练模式下可自适应输出高刚度、高阻尼导纳参数，实现高精度轨迹跟踪，在主动训练模式下的轨迹跟踪误差显著降低，人机交互力处于舒适区间，有效提升了下肢外骨骼控制的精准性、安全性与自适应能力。 |
| 关键词: 下肢外骨骼 导纳控制 深度强化学习 |
|
中图分类号:
文献标识码:
|
| 基金项目: 贵州省科技支撑计划项目(黔科合支撑[2023]一般179) |
|
| Admittance control of lower limb rehabilitation exoskeleton based on reinforcement learning |
|
hongtao, wuqinmu
|
Guizhou University
|
| Abstract: Traditional control strategies for lower limb exoskeletons struggle to dynamically adapt to different rehabilitation stages and pose safety risks in human-machine interaction. An improved SAC reinforcement learning adaptive admittance control algorithm with physical constraints is proposed to achieve collaborative optimization of exoskeleton trajectory tracking and human-machine interaction compliance. A constrained Markov decision process is introduced, and physical hard constraints are set using the Lagrangian multiplier method. The results show that the algorithm converges stably within 1000 training episodes, with the constraint violation cost eventually approaching 0. In passive training mode, the algorithm can adaptively output high stiffness and high damping admittance parameters to achieve high-precision trajectory tracking. In active training mode, the trajectory tracking error is significantly reduced, the human-machine interaction force is within a comfortable range, and the accuracy, safety, and adaptability of lower limb exoskeleton control are effectively improved. |
| Keywords: lower limb exoskeleton admittance control deep reinforcement learning |
