RoboLeague devlog #2: Training an AI to navigate an obstacle course with Unity ML agents

Follow the full discussion on Reddit.
Hello everyone! In my first devlog, I introduced you to RoboLeague - my open-source Rocket League clone for AI experiments and deep reinforcement learning research. Since the aerial controls and physics are identical to the real game I decided to do more experiments in the air. For this devlog, I managed to successfully train an agent with Unity ML agents to navigate one level of a popular obstacle course in Rocket League. To solve this problem the agent now has to use raycast observation, think of it as a 3D LIDAR sensor used in self-driving cars. Each ray measures the distance to the walls giving an agent the ability to sense the 3D environment akin to the echolocation of bats. In Unity, it looks like this.