Virtual Analysis of Vehicle Dynamics Limits using Reinforcement Learning

Vehicle dynamics is the key factor influencing the safety, comfort and maneuverability of a vehicle. Consequently, it is an essential area of research in automotive engineering. However, as vehicle performance, safety standards and the number of mechatronic systems, e.g. electronic stability control, continue to increase, assessing the limits of vehicle dynamics through traditional testing methods becomes increasingly challenging. Typically, simulated or real driving tests are conducted to analyze the limits. Nevertheless, finding the optimal set of test scenarios regarding the conflicting goals between coverage of the entire region of interest and minimal effort for testing is a time-consuming process when done manually. In this paper, a method based on Reinforcement Learning (RL) is presented for the automated exploration of the limits of vehicle dynamics. The developed co-simulation framework enables the seamless interaction between the RL agent acting as the virtual driver with the simulation environment. The aim is to uncover previously undiscovered test scenarios, thus creating the opportunity to assess and further improve the performance of vehicle stability systems.