CONTROLLING A DOUBLE-PENDULUM CRANE BY COMBINING REINFORCEMENT LEARNING AND CONVENTIONAL CONTROL

Controlling oscillation is vital for applications in which flexible systems are employed. Many existing control methods rely on knowledge of the system dynamics to mitigate unwanted vibration. However, model-free methods can also be employed to control vibration. One method for model-free control is reinforcement learning (RL). Although the RL agent does not require information about the system to learn a control policy, domain knowledge of dynamics and control can be used to augment the agent and aid in generating an effective control policy. This work analyzes the effectiveness of training RL controllers that operate in combination with conventional controllers. Agents were trained in simulation using a model of a small-scale double-pendulum crane. The effect of the conventional control component on training as well as sensitivity to modeling error are analyzed. Agent transferability is investigated by implementing the simulation-trained controllers on a physical small-scale double-pendulum crane.