An Expeditious and Expressive Vehicle Dynamics Model for Applications in Controls and Reinforcement Learning

Abstract

We present a Vehicle Model (VM) that has 17 degrees of freedom and includes nonlinear tire and powertrain subsystems. Implemented as a relatively small piece of C++ code, the model runs vehicle dynamics 2000 times faster than real time at a simulation time step of $1 \times 10^{-3}, \text {s}$ on a single core of a commodity CPU. When executed on the GPU, one can perform 300000 vehicle simulations in real-time. These properties make the model a good candidate for reinforcement learning, model predictive control, model predictive path integral control, path planning, state estimation, and traffic simulation tasks. The model is expressive in that it can capture the dynamics of vastly different vehicles. This is demonstrated by first calibrating the model to mimic the dynamics of a 1/ $6^{th}$ scale vehicle called the Autonomy Research Testbed (ART) vehicle, which has a mass of approximately 5.8 kg. Subsequently, the model is calibrated to mimic the dynamics of a heavy-duty High Mobility Multipurpose Wheeled Vehicle (HMMWV), which has a mass of 2097 kg. The Bayesian calibration process discussed can $(i)$ handle real-life measurement noise, and $(ii)$ provide model parameter probability distributions. The vehicle model, which is open source and freely available in a public repository, can also be imported into Python owing to SWIG wrapping.

Keywords

• Vehicle models
• Bayesian inference
• calibration
• state estimation
• control
• machine learning