Abstract
Software defined vehicles bring a different approach even into the development of vehicle motion control functions. In addition to the existing functions, many new functionalities are demanded, all while facing restricted development time and budget, however available information for starting a development might be very limited. When used correctly, Artificial Intelligence can help establish a new approach to satisfy these boundary conditions. AVL and the University of Surrey have been collaborating for years to explore how Deep Reinforcement Learning (DRL) can fit into this new approach. With this paper we present the way to a real-world implementation of an integrated chassis controller, developed by DRL in simulation and validated with experiments. The algorithm focuses both on traction control and suspension control. It is designed to consider time delays in the control chain, and it integrates a legacy code for semi active suspension control via a weight adaptation layer, also generated by DRL. The developed system is validated with a VW e-Golf in the AVL ZalaZone proving ground.