This paper proposes a multimodal perception-fusion-based control method for flexible exoskeleton robots to address challenges in human–robot collaboration and interaction accuracy. First, after analyzing the human–robot collaborative dynamics, the key control variables are calculated. Using these as a baseline, multimodal perception information is collected and preliminarily feature-extracted via sensor technology and lightweight deep learning models. Then, a deep learning architecture performs feature-level fusion of these multimodal signals to derive state vectors that reflect the lower-limb motion of the exoskeleton. Finally, a force-feedback-based impedance control principle combined with a proportional-integral-derivative (PID) scheme is adopted to design the controller, and the derived state vectors are used to adjust the controller parameters in real time, thereby achieving effective human–robot collaborative control. Experimental results show that when the proposed method is applied, the output human–robot interaction force curve is smooth and remains close to the zero axis with minor fluctuations. The maximum interaction force is 2.58 N, with a standard deviation of ∼0.75 N. Meanwhile, within the control duration, the root mean square error difference of hip joint tracking errors among multiple subjects remains stable between 0.42° and 0.56°, and the average absolute error remains stable between 0.34° and 0.45°. The root mean square error value difference of knee joint tracking error is basically stable between 1.36° and 1.51°, and the average absolute error value is basically stable between 1.09° and 1.20°. This indicates that it achieves high precision in human–machine collaborative motion trajectory tracking control and delivers good control performance.
J Q Li (Fri,) studied this question.