Deep Reinforcement Learning for a Humanoid Robot Basketball Player

摘要

Currently, the majority of research on humanoid robot basketball shooting focuses on traditional control methods. However, these methods primarily rely on human-robot interaction and fixed shooting patterns to control the robot’s shooting actions, resulting in limited autonomy for the robot. They often require extensive manual design and coding operations, and face challenges in adapting to different shooting scenarios. To address these problems, this paper applies deep reinforcement learning to the basketball shooting task for a humanoid robot. The task environment is based on the basketball shooting competition defined in the FIRA HuroCup. This paper uses the Double DQN algorithm to train the humanoid robot to master end-to-end basketball shooting skills, specifically: The robot takes RGB images captured by its own head camera as input, then decides to take one of three discrete actions, including turning left, turning right, and shooting. In the experimental section, we validate the effectiveness of our approach and conduct an analysis and discussion on the setup of important parameters that influence the experimental results.