Investigating Training an Agent to Drive a Vehicle in a Simulated Environment Using Reinforcement Learning

This thesis investigates if reinforcement learning (RL) is suitable for training an agent to drive a vehicle in a simulated environment that contains generated red and blue gates that the vehicle has to pass by using visual inputs from a camera. The generated goals should result in a map similar to a ski slalom. The vehicle then has to drive from the beginning to the end of the map and must pass all of the generated gates in the correct order as fast as possible. Additionally, the car and the environment are modeled as close as possible to reality. As model serves a constructed arena which is located at the ScaDS.AI living lab and as a vehicle a JetBot (jetbot.org) is used. However, all this ends up in a very complex challenge that the AI has to accomplish.

Counterpart