A TD3-Based Reinforcement Learning Algorithm with Curriculum Learning for Adaptive Yaw Stability Control in All-Wheel-Drive Electric Vehicles

A novel artificial intelligence-based approach for the direct yaw control (DYC) of an all-wheel drive (AWD) electric vehicle (EV) is proposed in this paper. To improve adaptability and ability to handle nonlinearities via continuous learning, the proposed algorithm is built upon a twin delayed deep deterministic policy gradient (TD3) reinforcement learning (RL) algorithm for the optimal torque distribution across four wheels of the vehicle. The proposed model-free torque vectoring algorithm performs based on the interaction of an agent with an environment to learn the optimal policy in a reward-driven manner and obtain the ability to dynamically adapt to varying conditions, such as different roads and vehicle speeds. Unlike conventional control methods that rely on precise system modeling and may struggle to adapt under varying conditions, no model of the vehicle is required in the proposed method. This work proposes a model-free RL-based controller with curriculum learning to train the strategy, where the model learns simpler tasks first, progressively increasing difficulty to enhance stability and convergence. A detailed reward function and well-structured actor-critic networks are devised, and the proposed algorithm is compared with a conventional model-based linear quadratic regulator (LQR) approach. A nonlinear model with 7 degrees of freedom is used to model the dynamic behavior of the vehicle in MATLAB/Simulink, and the results are further verified through the implementation of IPG CarMaker under realistic driving scenarios. The performance of the proposed algorithm is studied across different maneuvers, demonstrating reduced yaw rate error and sideslip angle, resulting in enhanced dynamic stability.