RocketDRL: a novel 3d environment for training reinforcement learning agents

Abstract

The development of autonomous agents capable of presenting more human-like behavior is currently driven by Deep Reinforcement Learning techniques. Deep Reinforcement Learning is an active field of research that is fueled by virtual environments usually inspired or borrowed from video games. Several works in the field are limited to playing classical control tasks, 2D environments, or outdated games. Therefore, most environments used in research are significantly different from those available in current trending 3D games. This paper introduces RocketDRL, a novel Deep Reinforcement Learning environment which supports mechanics for 3D games inspired by the popular “car football” game Rocket League. Besides the classical gameplay, we implemented three challenging minigames based on the mechanics from this title with an advanced simulation of physics with fine-grained car control: penalty shoot, free kick, and aerial shoot. Moreover, we also provide promising baseline results using Unity’s ML-Agents Toolkit, which is an easy way to train and evaluate the agents.