Control of an Underactuated Mechanical System with Reinforcement Learning Compensation
Rubén Hernández, Ramón García-Hernández, Francisco Jurado · 2023
In recent years, reinforcement learning (RL) algorithms have exhibited significant potential in controlling underactuated mechanical systems, offering a distinct approach from conventional control theory. However, RL agents trained in nominal environments may experience performance degradation when confronted with external disturbances. This work investigates the impact of combining the learned policy of a RL agent with a feedback linearization (FL) controller to compensate the nonlinearities and parametric uncertainties in order to stabilize the Furuta pendulum. This hybrid control scheme enhances the robustness of the acquired policy. Through numerical simulations, we compare the proposed scheme with an approach based on FL with adaptive neural compensation (ANC), revealing superior performance in the presence of external disturbances when RL-based compensation is used.