A Learning-Based Model Predictive Contouring Control for Vehicle Evasive Manoeuvres
Alberto Bertipaglia, Mohsen Alirezaei, Riender Happee, Barys Shyrokau
TL;DR
This work addresses evasive manoeuvres at the limit of handling by coupling a learning-based predictive framework with a Model Predictive Contouring Control (MPCC). It introduces a Learning-based MPCC (L-MPCC) that leverages a Student-T Process to predict and propagate uncertainties in tyre-force and yaw-rate mismatches, integrating these into the control cost to minimize model error and state uncertainty. The approach yields three main contributions: (i) direct reduction of tyre-model mismatches leading to higher allowable speeds, (ii) enhanced outlier robustness and measurement-informed uncertainty through the STP, and (iii) substantial improvements in stability, exemplified by a dramatic reduction in peak sideslip during evasive maneuvers. The results demonstrate improved safety margins and maneuverability in high-fidelity simulations, with potential benefits for real-time automated driving systems under uncertain tire dynamics.
Abstract
This paper presents a novel Learning-based Model Predictive Contouring Control (L-MPCC) algorithm for evasive manoeuvres at the limit of handling. The algorithm uses the Student-t Process (STP) to minimise model mismatches and uncertainties online. The proposed STP captures the mismatches between the prediction model and the measured lateral tyre forces and yaw rate. The mismatches correspond to the posterior means provided to the prediction model to improve its accuracy. Simultaneously, the posterior covariances are propagated to the vehicle lateral velocity and yaw rate along the prediction horizon. The STP posterior covariance directly depends on the variance of observed data, so its variance is more significant when the online measurements differ from the recorded ones in the training set and smaller in the opposite case. Thus, these covariances can be utilised in the L-MPCC's cost function to minimise the vehicle state uncertainties. In a high-fidelity simulation environment, we demonstrate that the proposed L-MPCC can successfully avoid obstacles, keeping the vehicle stable while driving a double lane change manoeuvre at a higher velocity than an MPCC without STP. Furthermore, the proposed controller yields a significantly lower peak sideslip angle, improving the vehicle's manoeuvrability compared to an L-MPCC with a Gaussian Process.