Error Compensation Strategies for Lower-Limb Rehabilitation Robots: A Staged Approach with MLP and Transformer Models

Lower-limb rehabilitation robots are valuable for gait training, but accurate joint motor angle tracking remains challenging due to various motion-related disturbances. This paper presents a staged joint-compensation strategy to improve control accuracy. The gait control process is partitioned into initiation, cyclic, and termination phases. A multilayer perceptron is employed during initiation and termination to predict and compensate for short-term aperiodic errors, while a Transformer-based sequence model combined with repetitive-control concepts is used in the cyclic phase to predict and correct periodic errors. Phase detection and safety-constraint mechanisms are integrated to ensure system stability and safety. Experiments are performed on a self-developed robotic platform with field-oriented control at the motor level, using a 165 cm, 60 kg dummy as the load. The proposed strategy substantially reduced joint-angle RMSE: left hip from 0.692° to 0.494° (28.6% reduction), right hip from 0.687° to 0.402° (41.5% reduction), left knee from 1.754° to 0.426° (75.7% reduction), and right knee from 1.667° to 0.461° (72.3% reduction). Ablation studies and repeated-trial statistical analyses further confirm the effectiveness of the approach. This study significantly reduces the gait trajectory tracking errors of joint actuators in a lower-limb rehabilitation robot, thereby providing a feasible and effective approach for the optimization of its control algorithm design.