Learning-Based Control of Humanoid Robots

This chapter will introduce the methods for motion control of humanoid robots based on deep reinforcement learning (DRL). Unlike traditional control methods, DRL-based methods require a large amount of data for policy optimization, so the entire learning and control process will focus on efficient learning and obtaining reliable data. This chapter will start with the basic DRL algorithms, first introducing several representative DRL-based methods to elucidate their structures and characteristics. Then, combined with specific humanoid robots and task scenarios, it will present the design ideas for DRL observations, action space, and rewards in common scenarios. Finally, based on the characteristics of DRL, it will explain the considerations during simulation, and establish simulation models and scenarios in line with specific contexts, introducing the construction of a complete DRL-based motion control method for humanoid robots.