In this talk, we provide recent results, and highlight the challenges, in reproducible software-in-the-loop simulation of cooperative experiments for fleets of autonomous and human-driven vehicles. Evaluation of the safety of vehicle behaviors requires both analysis and testing. As part of the testing approach for multi-vehicle behaviors, simulations must go beyond state data that describe position estimates and which provide an ability to set instantaneous acceleration or velocity values. High-fidelity simulation environments enable simulation of the motion of a vehicle through simulation of its dynamics, as well as the simulation or emulation of sensor behaviors: this permits software-in-the-loop evaluation of the behavior of candidate software that will be run in hardware. However, these simulation environments do not scale to even tens of vehicles, and simulations exhibit divergent trajectories from the same initial conditions, due to the inherent challenges of unsynchronized clocks as many different components are added to the system.
In this talk, we present an architecture to obtain more repeatable simulations via a layer of temporal logic and event-based design that is introduced at the component integration layer. The approach offloads the vehicle state dynamics evolution from high-fidelity simulators, removing unnecessary computational burden, and permitting a synchronized update rate for vehicle motion in order to purposefully reduce the real-time factor and permit more vehicles to take part in the simulation. This approximation of agent dynamics reduces system complexity, thus enhancing scalability of the simulation to dozens of autonomous vehicles with simulated sensors. We quantify the scalability and repeatability of our design via real-time factor deviation and mean-squared error of trajectory deviation.