Continuous Motion Prediction for Upper Limb Rehabilitation Robotics Based on Surface Electromyography: A Spatio-Temporal Attention and Lightweight TCN Approach

Surface electromyography (sEMG) has shown considerable promise for controlling upper limb rehabilitation exoskeletons. However, the significant inter-subject variability in sEM G signals poses a challenge for developing models that generalize effectively across individuals. A novel spatio-temporal attention mechanism integrated with a lightweight temporal convolutional network (STA - TCN) is proposed in this paper to address this challenge and enable subject-independent continuous motion prediction. By incorporating a dual-path (spatial and temporal) attention mechanism, the spatial (muscle channel) and temporal (movement phase) features related to sEMG signals are selectively emphasized, eliminating the need for subject-specific retraining. A Cable-driven Upper-Limb Rehabilitation Exoskeleton (CURE) is designed, for which continuous motion prediction was employed. An experiment with 6 participants demonstrated that STA- TCN outperforms traditional models in continuously predicting elbow angles, achieving lower root mean square error (RMSE) (14.77° vs 15.25°) and R2 values approaching the current highest results (0.91 vs 0.93) across users. The method also met real-time processing requirements with a latency of 1.5 ms per prediction, making it suitable for deployment in real-time rehabilitation robotics. The proposed method can provide an effective solution for personalized and adaptable robotic rehabilitation systems, enhancing motion prediction accuracy without dependency on individual calibration.