Chassis control (CC) plays a crucial role in ensuring vehicle performance, stability, and safety

through the individual or concurrent control of various chassis actuators. Over the past few

decades, deterministic approaches have been explored, such as proportional-integral-derivative

(PID) and model predictive control (MPC) strategies, demonstrating strong potential in

different fields, from traction to direct yaw moment control. However, these strategies are still

characterised by several challenges, primarily related to design complexity, tuning, estimation

requirements and real time capability. To tackle these challenges, reinforcement learning (RL)

agents have gained interest for their ability to learn control tasks by collecting data through

interaction with the environment, maximising a reward function over multiple iterations which

define the so-called training process. In this context, this project aimed of exploring the

capabilities and opportunities of RL for CC, with special focus on torque vectoring (TV) and

traction control (TC), in comparison with state-of-the-art controllers, e.g., MPC-based and

PID-based strategies. As further point, to investigate opportunities of adaptation on a real

vehicle, an innovative adaptation scheme is proposed to improve control robustness under

variations of friction conditions or vehicle parameters, such as mass and weight distribution.

The TC case study has been particularly explored, leading to an experimental campaign on a

Volkswagen eGolf across various friction conditions. Experimental results highlighted the

agents' superior performance compared to a state-of-the-art integral sliding mode controller.

The baseline RL-TC, i.e., not augmented with the adaptation scheme, increase the longitudinal

acceleration by ~40% while reducing the reference slip tracking error by ~20%. Furthermore,

the experimental campaign highlighted the adaptation framework’s value in improving

robustness to vehicle and friction changes, ensuring the same level of acceleration of the

baseline RL controller, while halving the tracking error of the reference slip.