Model predictive control (MPC) has proven useful in enabling safe and optimal motion planning for autonomous vehicles. In this paper, we investigate how to achieve MPC-based motion planning when a neural state-space model represents the vehicle dynamics. As the neural state-space model will lead to highly complex, nonlinear and nonconvex optimization landscapes, mainstream gradient-based MPC methods will struggle to provide viable solutions due to heavy computational load. In a departure, we propose the idea of model predictive inferential control (MPIC), which seeks to infer the best control decisions from the control objectives and constraints. Following this idea, we convert the MPC problem for motion planning into a Bayesian state estimation problem. Then, we develop a new implicit particle filtering/smoothing approach to perform the estimation. This approach is implemented as banks of unscented Kalman filters/smoothers and offers high sampling efficiency, fast computation, and estimation accuracy. We evaluate the MPIC approach through a simulation study of autonomous driving in different scenarios, along with an exhaustive comparison with gradient-based MPC. The simulation results show that the MPIC approach has considerable computational efficiency despite complex neural network architectures and the capability to solve large-scale MPC problems for neural state-space models.

翻译：模型预测控制（MPC）已被证明在实现自动驾驶车辆安全与最优运动规划方面具有重要价值。本文研究了当车辆动力学由神经状态空间模型表示时，如何实现基于MPC的运动规划。由于神经状态空间模型将导致高度复杂、非线性且非凸的优化空间，主流的基于梯度的MPC方法因计算负荷过重而难以提供可行解。为此，我们提出模型预测推理控制（MPIC）的思想，旨在从控制目标与约束中推断最优控制决策。基于这一思想，我们将运动规划的MPC问题转化为贝叶斯状态估计问题。随后，我们开发了一种新的隐式粒子滤波/平滑方法进行估计。该方法通过多组无迹卡尔曼滤波器/平滑器实现，具有高采样效率、快速计算与精确估计的特点。我们通过不同场景下的自动驾驶仿真研究，并与基于梯度的MPC方法进行详尽对比，评估了MPIC方法的性能。仿真结果表明，即使面对复杂的神经网络结构，MPIC方法仍具有显著的计算效率，并能解决神经状态空间模型的大规模MPC问题。