Learning to Play Precision Ball Sports from scratch: a Deep Reinforcement Learning Approach

摘要

Over the last years, robotics has increased its interest in learning human-like behaviors and activities. One of the most common actions searched, as well as one of the most fun to replicate, is the ability to play sports. This has been made possible with the steady increase of automated learning, encouraged by the tremendous developments in computational power and improved reinforcement learning (RL) algorithms.This paper implements a beginner Robot player for precision ball sports like boccia and bocce. A new simulated environment (PrecisionBall) is created, and a seven degree-of-freedom (DoF) robotic arm, is able to learn from scratch how to win the game and throw different types of balls towards the goal (the jack), using deep reinforcement learning. The environment is compliant with OpenAI Gym, using the MuJoCo realistic physics engine for a realistic simulation. A brief comparison of the convergence of different RL algorithms is performed. Several ball weights and various types of materials correspondent to bocce and boccia are tested, as well as different friction coefficients. Results show that the robot achieves a maximum success rate of 92.7% and mean of 75.7% for the best case. While learning to play these sports with the DDPG+HER algorithm, the robotic agent acquired some relevant skills that allowed it to win.