Utilizing Reinforcement Learning to Improve Gait Training Performance and Development of Gravity-Compensating Control System for RoboREHAB Device

Abstract As a patient gets older, the chances of sustaining a motor disability increase. Such disabilities impair the patient’s ability to perform daily activities and undergo rehabilitation to regain mobility. Exoskeleton devices and aqua therapy have been used to make this process easier. However, most exoskeletons do not replicate the natural gait pattern accurately and there does not exist a lower-limb rehabilitation device that provides gait training with zero-gravity assistance. The objective for this study is to design two control systems that govern motor control for the natural gait pattern using a redesigned leg assembly and create a weight-compensation control system for a traditional leg assembly. Reinforcement learning (RL) was used in Simulink to develop the gait pattern control system, and a physical prototype was created to test the zero-gravity control system using Arduino. The RL control system with the redesigned leg showed a X- and Y-deviation of 45.622 mm and 17.164 mm at the knee joint compared to the original two-segment design, which achieved an X- and Y-deviation of 133.333 mm and 38.832 mm at the knee joint. Zero-gravity was achieved via DC Motors for the knee and foot sections, but the motors were not strong enough to support the load at the hip joint. Using Servo Motors, a sit-to-stand movement was created, with only a slight reduction in the range of motion of the hip due to the 5V motors limitations. Finally, a walking motion was created using Servo Motors, achieving another method of zero-gravity. The zero-gravity controller worked for the walking and sit-to-stand transfer but the hip joint needs further improvement.