Deep Reinforcement Learning for a Mechanical System under Friction Effect

When a reinforcement learning (RL) agent is trained in a nominal environment and then applied to an uncertain system, the control policy obtained could fails when the environment is then perturbed by external disturbances and friction forces. In order to verify this fact, in this work two RL agents are designed; the first one is trained when considering a nominal model and the last one is under an environment in presence of static friction with readjustment of the reward function. A PD controller is added to each agent in order to improve the previously trained policy. This hybrid controller improves the robustness of the trained policy. Numerical simulations from considering an underactuated mechanical system model under friction effects show the feasibility of our proposal.