Abstract
Most of the research on vehicle stability control systems has dealt with feedback approaches, e.g., based on proportional-integral (PI), linear quadratic and model predictive control, which have been developed to generate direct yaw moments and longitudinal tire force distributions to compensate for large yaw rate errors and/or sideslip angle magnitudes, conditions in which the driver steering inputs are not able to significantly influence the cornering response. However, these strategies are still characterized by multiple challenges, such as the design complexity related to the necessity of considering tire nonlinearities and actuator-related dynamics and delays. To this purpose, in the last few years, artificial intelligence solutions have been proposed, among which reinforcement learning (RL) has shown promising results, thanks to its ability to solve the control task by learning through trial and error. In this field, only a few studies have evaluated RL applications for torque-vectoring control, with none of them considering the effect of the path tracking control layer in the context of automated vehicles. In this research, a real-time capable RL torque-vectoring controller for an automated vehicle is developed and assessed in obstacle avoidance maneuvers (according to the ISO 3888 standard), through an experimentally validated high-fidelity vehicle simulation model, including an automated driving control layer, as well as realistic tire behavior and actuation dynamics. In comparison with a rule-based PI algorithm, the results highlight the RL capability of flexibly tailoring the vehicle response depending on the selected reward function. The proposed RL agent’s ability to learn control tasks through high-fidelity simulations could accelerate advanced control strategy design, and support benchmarking activities, by demonstrating optimal solutions for specific scenarios.