Deep Dynamics: Vehicle Dynamics Modeling with a Physics-Informed Neural Network for Autonomous Racing

Author: ORCID icon
Chrosniak, John, Computer Science - School of Engineering and Applied Science, University of Virginia
Behl, Madhur, EN-Comp Science Dept, University of Virginia
Kuo, Yen-Ling, EN-Comp Science Dept, University of Virginia
Bezzo, Nicola, EN-SIE, University of Virginia

Autonomous racing is a critical research area for improving the safety of autonomous driving. By exposing autonomous vehicles to high speeds (>280 kph) and accelerations, autonomous racing competitions offer an exciting platform to stress test the capabilities of autonomous driving software and improve the algorithms needed for advanced collision avoidance on public roads. The conditions observed in a racing environment present significant challenges in vehicle dynamics modeling, such as balancing model precision and computational efficiency under extreme dynamic loads, where minor errors in modeling can have severe consequences. Physics-based models require time-intensive and cost-prohibitive testing environments to accurately identify model coefficients that capture the vehicle's underlying dynamics, with frequent recalculations needed due to varying vehicle configurations, track conditions, and wear of vehicle components like tires and brakes. Conversely, although easier to identify, purely data-driven approaches using deep neural networks (DNNs) in place of a physics-based model do not generalize well and provide no guarantees predictions will make physical sense. This thesis introduces Deep Dynamics, a physics-informed neural network (PINN) for vehicle dynamics modeling of an autonomous racecar that combines the usage of a DNN for physics coefficient estimation with dynamical equations to accurately predict vehicle states at high speeds. With little prior knowledge of the vehicle needed, Deep Dynamics indirectly learns the complex dynamics of a racecar solely through historical observations of the vehicle's states and control inputs. The inclusion of a novel Physics Guard layer constrains internal coefficient estimates to always remain within their nominal physical ranges, guaranteeing a valid physics model is produced. When trained using data collected from a physics-based simulator and a full-scale autonomous Indy racecar, Deep Dynamics displays superior open-loop performance for single timestep and forward horizon predictions in comparison to a previously introduced PINN vehicle dynamics model. An analysis of the internally estimated coefficients confirms that this improved accuracy is attributed to a better understanding of the vehicle's physical properties. Closed-loop testing in the physics-based simulator also showcases Deep Dynamics' ability to safely operate an autonomous racecar at its handling limits, exhibiting faster lap times and higher average speeds while remaining within the boundaries of the racetrack. Based on these assessments, Deep Dynamics offers a promising approach for modeling racecar vehicle dynamics.

MS (Master of Science)
Autonomous Racing, Vehicle Dynamics Modeling, Physics-Informed Neural Networks, Robot Learning, Model Learning for Control
All rights reserved (no additional license for public reuse)
Issued Date: