Driving Policy Transfer via Modularity and Abstraction

Abstract

Driving Policy Transfer via Modularity and Abstraction

Summary ➢We transfer driving policies from simulation to reality via modularity and abstraction. ➢The driving policy is encapsulated such that it is not directly exposed to raw perceptual input or low-level vehicle dynamics. ➢We evaluate our approach in simulated urban environments and in various real-world conditions in two different continents.

Simulation ➢We use CARLA, an open-source simulator for urban driving. ➢The simulator provides access to sensor data from the ego-vehicle, as well as detailed privileged information about the ego-vehicle and the environment. ➢CARLA provides access to two towns: Town 1 and Town 2 which differ in their layout, size, and visual style. ➢CARLA also provides multiple environmental conditions (combinations of weather and lighting). ➢We use two of these in our experiments: clear daytime and cloudy daytime after rain. ➢The two towns and environmental conditions used in our experiments are illustrated on the right.

Physical World ➢We use a modified 1/5 scale Traxxas Maxx truck as vehicle. ➢At runtime, given an image, the onboard computer predicts the waypoints and uses a PID controller to convert them to low-level control commands. ➢While the car is driving, the driving policy can be guided by high-level command inputs through a switch on the remote control.