Deep reinforcement learning of robotic prosthesis for gait symmetry in trans-femoral amputated patients

This work proposes a novel control methodology to achieve gait symmetry in trans-femoral amputated patients with prostheses. The proposed approach allows to overcome the limits of classical model-based control strategies by introducing a Deep Reinforcement Learning (DRL) method trained ad hoc for generating the velocity control signals fed into the active lower-limb robotic prosthesis. More specifically, the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm is used to concurrently learn a Q-function and the best policy. The proposal has the advantages of being model-free and capable of adapting to different walking velocities, just requiring few measurements and without the need to online re-tune the control parameters when the human motions change. The proposed model-free approach has been tested in a realistic scenario simulated in the CoppeliaSim environment relying on gait patterns retrieved experimentally by means of markers placed on a human subject.