Leveraging control priors in deep reinforcement learning for robotics

摘要

This thesis proposes a series of hybrid approaches to robot control that combine classical control methods and deep reinforcement learning (RL). Classical control methods work well in structured environments, but struggle in unstructured and stochastic situations. RL has the potential to learn complex controllers through trial and error, however, current methods are sample-inefficient and unsafe. The proposed hybrid approaches combine the strengths of both systems, resulting in efficient, reliable, and dexterous decision-making systems for real-world robotics. This is demonstrated through a sequence of four publications, showing that the hybrid approaches outperform either system operating independently.