Repeatable & Scalable Multi-Vehicle Simulation with Offloaded Dynamics using Federated Modeling

In this paper, we present a method to perform multi-vehicle simulation of autonomous systems that improves the repeatability of robotics simulations and can improve the scale of such simulations for dynamically complex devices such as autonomous vehicles (AV). Current approaches to simulation of multi-component AV typically infer the kinematics or dynamics through the rigid-body motion that uses joint angles and shapes. Such simulations encounter challenges for simulated AV, as the methods to discretize the behavior are prone to error accumulated over time and are computationally intensive – frequently resulting in chaotic behavior. The accumulated error results in a lack of repeatability of the simulation results. Further, when simulating multiple AVs, simulations typically fail to scale to tens of vehicles, even when slowed to permit more accurate results as the state evolves. This paper provides an architecture for improving the repeatability of simulations using federated modeling and state synchronization through a director. The method consists of replacing the inverse kinematic vehicle models with computational models of their dynamics, offloading dynamics from the physics engine for state evolution, synchronizing vehicle updates using a director, and performing the simulation at slower than real-time if needed. Our method reduces the error of trajectory deviation during repeated simulations by at least threefold. An implementation of the results of the work is presented through a Robot Operating System (ROS) package.