Learning-based model predictive control with moving horizon state estimation for autonomous racing

TL;DR

This work tackles autonomous racing by integrating a real-time nonlinear model predictive controller (NMPC) with a moving horizon estimator (MHE) and an online Gaussian-process (GP) learning extension to compensate model mismatch. An offline high-fidelity NMPC-based planner generates a time-optimal race line under vehicle and track constraints, which the real-time, low-order NMPC then tracks with state estimates provided by MHE. The GP corrections refine predictions in aggressive maneuvers, enabling higher speeds and tighter tracking on challenging tracks, while preserving real-time feasibility on standard hardware. The approach is validated on two tracks, showing improved tracking accuracy and speed over a baseline MPC, with real-time computation demonstrated at approximately 130 Hz on commodity hardware.

Abstract

This paper addresses autonomous racing by introducing a real-time nonlinear model predictive controller (NMPC) coupled with a moving horizon estimator (MHE). The racing problem is solved by an NMPC-based off-line trajectory planner that computes the best trajectory while considering the physical limits of the vehicle and circuit constraints. The developed controller is further enhanced with a learning extension based on Gaussian process regression that improves model predictions. The proposed control, estimation, and planning schemes are evaluated on two different race tracks. Code can be found here: https://github.com/yassinekebbati/GP_Learning-based_MPC_with_MHE

Figures (19)

• Figure 1: Bicycle model.
• Figure 2: Moving horizon estimation process.
• Figure 3: Overall control approach.
• Figure 4: Control signals (measurement vs true value).
• Figure 5: Vehicle states estimation with MHE.