Deep Learning-Based Real-Time Model Predictive Motion Control for Distributed Drive Electric Vehicles Under Extreme Conditions

The motion control of distributed drive electric vehicles (DDEVs) under extreme operating conditions is significantly affected by the strong nonlinearities in vehicle dynamics. A novel hierarchical control architecture is proposed to enhance the control performance under such conditions. The upper-level deep learning-based model predictive motion controller determines the front wheel steering angle, longitudinal force, and additional yaw moment based on the vehicle's current state, target trajectory, and sideslip angle at the center of gravity (CG) constraints. It integrates low-fidelity, interpretable physical models with neural networks (NNs), leveraging automatic differentiation to backpropagate the nonlinear model predictive control (NMPC) loss function through a differentiable nonlinear dynamics model to obtain policy gradients, while using tailored NN output layers and an augmented Lagrangian method to ensure the satisfaction of control and state constraints. The lower-level torque distributor is constructed to prioritize vehicle stability by minimizing tire adhesion utilization. High-fidelity co-simulations and hardware-in-the-loop (HIL) experiments including an embedded onboard controller demonstrate the effectiveness of the proposed hierarchical control scheme in terms of real-time computation and closed-loop control performance. The real-vehicle test results also demonstrate that our algorithm can effectively handle state constraints and ensure vehicle stability under extreme conditions.