Methods to Cross the simulation-to-reality gap on a Autonomous Driving Model Car

The Reinforcement Learning (RL) method for training a self-driving model car in a simulated environment (Unity) was investigated in a master’s thesis.
A parkour set with red and blue pillars is driven by a model car in a virtual arena. A front camera is mounted on the model car.
The image sequences are processed with Computer Based Vision (CV). They serve as input data for the RL model.
The environment as well as the model car (jetbot.org) have been modelled as realistically as possible in Unity.
The task of the bachelor thesis is to transfer and evaluate the RL model trained in the simulation to the real model car.
In particular, gaps between simulation and reality are to be identified and solutions developed to close them.

Counterpart