Gait Phase-Free Neural Network-Based Controller for Cable-Driven Assistive Suits

Human assistive systems such as exoskeleton robots and soft suits have been introduced to aid rehabilitation and prevent fatigue by assisting skeletal muscles. The performance of these systems depends on the timing and amount of force generation by the controllers. Designing phase-dependent controllers is challenging due to variations in human movement patterns. Recently, deep reinforcement learning (DRL) has shown potential in resolving control problems using deep neural network-based controllers. The objective of this study was to develop a phase-free controller for a cable-driven assistive suit using forward dynamics simulation. A three-dimensional full-body human model and a cable-driven suit model were prepared in the MuJoCo. A human controller and a suit controller were trained using DRL. To recognize gait states, maintaining a minimum tensional force of 10 N was included in the reward function during suit controller training. The mechanical work of the hip decreased when the assistive forces of the cable-driven suit were applied. The positive work during walking was 0.53 J/kg with the suit on and 0.95 J/kg with the suit off. This study demonstrated the effectiveness of a neural network-based suit controller in assisting hip joint power during the stance phase of walking without relying on a gait phase detection method.