FLUIDS: A First-Order Lightweight Urban Intersection Driving Simulator

Abstract

To facilitate automation of urban driving, we present an efficient, lightweight, open-source, first-order simulator with associated graphical display and algorithmic supervisors. FLUIDS can efficiently simulate traffic intersections with varying state configurations for the training and evaluation of learning algorithms. FLUIDS supports an image-based birdseye state space and a lower dimensional quasi-LIDAR representation. FLUIDS additionally provides algorithmic supervisors for simulating realistic behavior of pedestrians and cars in the environment. FLUIDS generates data in parallel at 4000 state-action pairs per minute and evaluates in parallel an imitation learned policy at 20K evaluations per minute. A velocity controller for avoiding collisions and obeying traffic laws using imitation learning was learned from demonstration. We additionally demonstrate the flexibility of FLUIDS by reporting an extensive sensitivity analysis of the learned model to simulation parameters. FLUIDS 1.0 is available at https://berkeleyautomation.github.io/Urban_Driving_Simulator/.