Deep Reinforcement and Imitation Learning for Self-driving Tasks

Abstract

In this paper we train four different deep reinforcement and imitation learning agents on two self-driving tasks. The environment is a driving simulator in which the car is virtually equipped with a monocular RGB-D camera in the windshield, has a sensor in the speedometer and actuators in the brakes, accelerator and steering wheel. In the imitation learning framework, the human expert sees a photorealistic road and the speedometer, and acts with pedals and steering wheel. To be efficient, the state is a representation in the feature space extracted from the RGB images with a variational autoencoder, which is trained before running any simulation with a loss that attempts to reconstruct three images, the same RGB input, the depth image and the segmented image.