A Data-Driven Adaptive Control Algorithm for Lower Limb Rehabilitation Exoskeletons Based on Transformer

Lower-limb active exoskeletons are a promising technology in post-stroke rehabilitation training which is capable of providing extra torques to compensate for muscle weakness and promote functional recovery. The key is accurately predicting the assistive torque required to achieve personalized and timely torque output for the user's future movements. This paper proposes an end-to-end control method based on temporal prediction models, using simulation-generated data to train deep learning models, thereby enabling control algorithm modeling without real hardware.A musculoskeletal simulation platform is used to construct motion data under hemiplegic conditions based on motion capture data from healthy individuals. Inverse dynamics is then employed to compute the assistive torques that the exoskeleton should apply. A temporal dataset is subsequently built by integrating IMU sensor inputs, and a predictive model based on the Transformer architecture is trained to predict the exoskeleton control torque output over the next 100 milliseconds, achieving dynamic, real-time, and personalized assistive control.Experimental results show that this method achieves a relative error of less than 5% across three typical gait tasks, effectively predicting assistive torque during lower limb rehabilitation. It outperforms traditional control methods in terms of accuracy, smoothness, and trajectory error.