In this project, a reinforcement learning algorithm based on a neural network is designed for the autonomous drifting of model cars. The goal is to implement the learned controller on a prototype car, which is then able to drift along a circle, a figure 8, and finally an arbitrary track. The learning process is divided into to parts: In phase 1, a controller is pre-learned using a high-fidelity simulation of the car and the computational capacities of the available KI-Lab servers. Based on this result, in phase 2, a refinement of the weights takes place on the model car, using an less powerful embedded computing platform (NVIDIA Jetson Nano).

Team & Contact

• Prof. Dr. Georg Schildbach (georg.schildbach@uni-luebeck.de)
• Fabian Domberg