Impacts of Model Fidelity on Trajectory Optimization for Autonomous Vehicles in Extreme Maneuvers

In this paper, we quantify the impacts of model fidelity on the effectiveness of trajectory optimization for autonomous vehicles when driving at the limits of friction through experiments with a full-size vehicle. Models ranging from a double-track model with lateral and longitudinal load transfer dynamics to a simple point-mass model are used in combination with direct numerical optimization to generate optimal trajectories subject to the limits imposed by each model. The effectiveness of each model for trajectory planning is evaluated by testing the trajectories on an automated vehicle across friction conditions ranging from ice to dry asphalt. Comparisons between the outright performance of the car and the car's ability to track the optimal trajectory are made across the various models. The tests reveal that the advantage of more complex models is less that they better predict the vehicle's behavior, but that they provide a more nuanced view of the vehicle's limits and guidance on the proper coordination of the various actuators on the vehicle in order to make most efficient use of the available tire friction.