A Reinforcement Learning Controller for the Swing-Up of the Furuta Pendulum

In this paper, a nonlinear control strategy is developed by applying the Reinforcement Learning (RL) approach to control engineering tasks. To this end, the swing-up problem of the Furuta pendulum is solved as a benchmark example considering the presence of dry friction as a function of the instantaneous reaction forces. In the paper, a detailed description of the mechanical system is provided, including the equations of motion and the reward function used in the control algorithm based on continuous and sparse signals. The performance of the deep deterministic policy gradient algorithm in the proposed environment is also evaluated by means of numerical experiments.