Motion Intention Recognition and DDPG-Based Adaptive Impedance Control for a Robotic Upper-Limb Exoskeleton

This paper presents a robotic upper-limb rehabilitation exoskeleton for individuals with upper-limb motor impairments in the middle-to-late stages of rehabilitation. The exoskeleton features a metamorphic mechanical architecture capable of switching among four metamorphic configurations: shoulder adduction/abduction (SA/A), shoulder flexion/extension (SF/E), elbow flexion/extension (EF/E), and forearm pronation/supination (FP/S). For hierarchical assistance, a termite life cycle optimizer-tuned support vector machine (TLCO-SVM) is developed for metamorphic-configuration recognition, and a TLCO-optimized long short-term memory network (TLCO-LSTM) is proposed to predict the desired joint-angle trajectory. Based on the recognized configuration and the predicted desired joint-angle trajectory, a deep deterministic policy gradient-based adaptive impedance controller is developed to generate assistive torques and support compliant physical human-robot interaction. Experiments were conducted to evaluate the proposed recognition, prediction, and control framework. The TLCO-SVM achieves an average classification accuracy of 98.10%. The TLCO-LSTM achieves root mean square errors (RMSEs) of 2.71° (SA/A), 2.41° (SF/E), 3.47° (EF/E), and 5.19° (FP/S), respectively. Assistive-torque tracking RMSEs are 0.2497 Nm, 0.2252 Nm, 0.1130 Nm, and 0.3423 Nm for SA/A, SF/E, EF/E, and FP/S, respectively.