Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Model Predictive Inferential Control of Neural State-Space Models for Autonomous Vehicle Motion Planning

Iman Askari, Ali Vaziri, Xuemin Tu, Shen Zeng, Huazhen Fang

TL;DR

This work addresses motion planning for autonomous vehicles when dynamics are modeled by neural state-space representations, where gradient-based MPC is hindered by nonconvexity and computation. It introduces Model Predictive Inferential Control (MPIC), reframing MPC as Bayesian state estimation and solving with MPIC-X, a fast, scalable algorithm built from banks of unscented Kalman filters to perform implicit particle filtering and smoothing. The approach establishes an equivalence between MPC and a horizon-based Bayesian estimation problem, and uses Kalman-IPF/Kalman-IPS together with the unscented transform to infer optimal plans with far fewer particles, achieving orders-of-magnitude speedups over traditional solvers while maintaining competitive costs. Extensive simulations on highway driving and a real-world DevBot 2.0 vehicle demonstrate MPIC-X’s robustness to NSS complexity, long horizons, and varying neural architectures, highlighting potential for practical deployment in ML-driven autonomous navigation. Overall, MPIC provides a principled estimation-based route for MPC with neural dynamics, enabling safe, efficient, and scalable motion planning in complex driving scenarios and beyond.

Abstract

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.

Model Predictive Inferential Control of Neural State-Space Models for Autonomous Vehicle Motion Planning

TL;DR

This work addresses motion planning for autonomous vehicles when dynamics are modeled by neural state-space representations, where gradient-based MPC is hindered by nonconvexity and computation. It introduces Model Predictive Inferential Control (MPIC), reframing MPC as Bayesian state estimation and solving with MPIC-X, a fast, scalable algorithm built from banks of unscented Kalman filters to perform implicit particle filtering and smoothing. The approach establishes an equivalence between MPC and a horizon-based Bayesian estimation problem, and uses Kalman-IPF/Kalman-IPS together with the unscented transform to infer optimal plans with far fewer particles, achieving orders-of-magnitude speedups over traditional solvers while maintaining competitive costs. Extensive simulations on highway driving and a real-world DevBot 2.0 vehicle demonstrate MPIC-X’s robustness to NSS complexity, long horizons, and varying neural architectures, highlighting potential for practical deployment in ML-driven autonomous navigation. Overall, MPIC provides a principled estimation-based route for MPC with neural dynamics, enabling safe, efficient, and scalable motion planning in complex driving scenarios and beyond.

Abstract

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.
Paper Structure (25 sections, 1 theorem, 60 equations, 14 figures, 2 tables, 5 algorithms)

This paper contains 25 sections, 1 theorem, 60 equations, 14 figures, 2 tables, 5 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Autonomous Vehicle Motion Planning
  4. Optimal Control via Estimation
  5. MPC of Machine Learning Dynamic Models
  6. Highlights of Differences of the Study
  7. Overview of Motion Planning
  8. Neural State-Space Vehicle Modeling
  9. Driving Requirements and Objectives
  10. Motion Planning Synthesis
  11. Model Predictive Inferential Control
  12. MPC Through the Lens of State Estimation
  13. MPIC Based on Bayesian State Estimation
  14. Implicit Particle Filtering & Smoothing for MPIC
  15. Implicit Importance Sampling
  16. ...and 10 more sections

Key Result

Theorem 1

Assume that $\bar{\bm w}_{k:k+H}$ and $\bar{\bm v}_{k:k+H}$ are mutually independent white Gaussian noise processes with where Then, the problems in MHE-problem and MAP-problem will have the same optima.

Figures (14)

  • Figure 1: Neural state-space model for vehicle dynamics.
  • Figure 2: Trajectories of the EV (in red) and OVs (in green and blue) in the overtaking scenario. The color change from light to dark indicates the passage of the driving time.
  • Figure 3: Simulation results for the overtaking scenario: (a) the acceleration and incremental acceleration profiles in solid curves, with respective bounds in dashed lines; (b) the steering and incremental steering profiles in solid lines, with their respective bounds in dashed lines; (c) distance between the EV and OVs, with the dashed-line safe margin.
  • Figure 4: Computational time comparison between MPIC-X algorithm and gradient-based MPC for different networks, horizon lengths, and particle numbers: (a) Net-1; (b) Net-2; (c) Net-3.
  • Figure 5: Cost performance comparison between the MPIC-X algorithm with ten particles and the gradient-based MPC for Net-1 at different horizon lengths: (a) $H=10$; (b) $H=20$; (c) $H=40$; (d) $H=60$. The red-shaded region represents $\pm 2\sigma$ bounds for ten Monte Carlo runs of the MPIC-X algorithm.
  • ...and 9 more figures

Theorems & Definitions (4)

  • Remark 1
  • Theorem 1
  • proof
  • Remark 2