MPC for Off-Road Vehicle Trajectory Tracking

Abstract

This thesis explores the trajectory tracking control of autonomous off-road vehicles. The research aims to create a suitable trajectory tracking controller for use in industries such as mining, agriculture, and material transport, where enhancing safety and efficiency is paramount. With these applications comes the need to translate across varying off-road conditions. The complexity of off-road environments poses significant challenges, including varying bank and inclination angles, changing traction conditions, vehicle payloads, and complex terrain-tire dynamics. Furthermore, the need for real-time performance mandates that the controller be efficient, as traditional Non-linear Model Predictive Control (NL-MPC) is too computationally intensive for practical and cost efficient use.

To tackle these challenges, a coupled controller for longitudinal and lateral control has been developed using a dual-track vehicle with a linear tire model. This physics-based approach incorporates road angles into the formulation to account for significant bank and inclination angles. Simulation results in off-road scenarios indicate that Road Angle Model Predictive Control (RA-MPC) shows potential for improving trajectory tracking. However, in practical applications, accurately estimating these angles remains difficult due to the varying planes on which the tires operate. This direct modelling approach also limits the generalizability of the controller in off-road conditions as other sources of unmodelled dynamics are ignored.

To enhance the controller's performance in a broader way, a separate method of compensating for off-road modelling complexities through the use learning methods is explored. Gaussian Process Regression (GPR) is employed to improve tracking performance through data-driven modelling of complex off-road dynamics. While this thesis focuses on the integration and inital proof of concept using learning to augment the MPC formulation, the results demonstrate that GPR-MPC can effectively compensate for path inclination and bank angles as well as other sources of unmodelled dynamics. Notably, GPR-MPC excels in low friction (low mu) scenarios where there are significant parameter mismatches in the physics-driven MPC formulation.