Reinforcement learning based control for torque allocation in electric vehicles: a preliminary analysis

This article conducts a preliminary exploration of an innovative Reinforcement Learning RL-based control system applied to the Torque Allocation problem in a fully electric All-Wheel-Drive vehicle. The investigation delves into the untapped degrees of freedom in four-motor Electric Vehicles beyond total torque request and Torque Vectoring bias. Utilizing a Deep Deterministic Policy Gradient (DDPG) agent, the RL architecture is implemented within MATLAB/Simulink, incorporating co-simulation with VI-CarRealTime for vehicle dynamics. Comparative analysis against reference Torque Allocation strategies (open differential, FWD, RWD) is performed, assessing key performance factors in the reward function. The most successful RL system, trained with Second Order Sliding Mode Suboptimal torque vectoring algorithm, surpasses the average performance of reference strategies. Nevertheless, challenges such as marginal advantages, repeatability issues, prolonged training durations, and a lack of interpretability are noted.