Infinite-time model predictive motion control for autonomous driving considering delay in resource-constrained computing platforms

Abstract

Although optimization-based methods have become mainstream in autonomous driving motion control, their high computational complexity and significant memory usage still limit their deployment on mass-production platforms with limited resources. Currently, mainstream Model Predictive Control (MPC) solvers often fail to fully exploit the special structure of MPC and require large-scale matrix KKT decomposition, leading to high computational complexity. In contrast, the Linear Quadratic Regulator (LQR) requires less computation and maintains a high stability margin, but it cannot consider controller input delays and time-varying references. Motivated by the dynamic programming derivation of the LQR, this paper leverages the specific structure of the MPC problem to derive an infinite horizon solution method. This approach explicitly handles steering delays and time-varying references while eliminating the need for large-scale KKT matrix decomposition, thereby significantly improving real-time performance. In summary, the proposed MPC strategy maintains the computational efficiency and stability of the LQR. Furthermore, the algorithm accommodates input delays and time-varying references to significantly improve control accuracy and stability in delayed scenarios. Ultimately, the designed algorithm was successfully implemented with a maximum single-step runtime of $0.4\mathrm{ms}$ on a controller with a main frequency of $300\mathrm{MHz}$ and a Flash capacity of $5\mathrm{MB}$ . Additionally, considering actuator delays effectively improved the controller’s stability and control accuracy.