Learning-based real-time model predictive tracking control for autonomous vehicles with path-pattern adaptability

Abstract

Real-time implementation of nonlinear model predictive control (NMPC) for path tracking in autonomous vehicles (AVs) remains challenging due to the computational complexity of solving high-dimensional, nonlinear optimization problems under dynamic constraints. To address this issue, this paper proposes a novel end-to-end training framework, called the neural network optimizer (NN Optimizer), which significantly reduces the computational burden of NMPC, enabling real-time implementation. Specifically, the NN Optimizer integrates low-fidelity, interpretable physical models with neural networks (NNs), utilizing automatic differentiation to backpropagate the NMPC loss function through a differentiable dynamics model to obtain policy gradients. As opposed to traditional imitation learning (IL)-based approaches, usually reliant on expert knowledge, the proposed method does not rely on extensive labeled data, offering greater interpretability and path-pattern adaptability. During online implementation, it just involves simple function evaluation, avoiding the computation of gradient information and multiple iterations. In two validation scenarios during co-simulation, the NN Optimizer achieves a solving speed over 70 times faster than interior point optimizer (IPOPT), while improving closed-loop control performance by more than 10% compared to IL. In real-vehicle tests, the NN optimizer outperforms the linear quadratic regulator (LQR), achieving improvements in control performance of over 30% in the weave scenario and over 60% in the double lane change (DLC) scenario.