Deep reinforcement learning based autonomous driving with expert demonstrations

Abstract

Establishing a scalable and stable driving system from simple sensor modules in an autonomous driving system is essential for popularizing and upgrading autonomous driving. To this end, the End-to-End method constructs the process of detection, decision, and control from the sensor module with a deep neural network, learning the optimal relationship between sensor measurements and control values on its own and establishing a driving policy. The goal of this paper is to propose a self-driving technology that extracts semantic low-dimensional information from front camera sensor and navigation information, and uses deep reinforcement learning induced by expert demonstration to control longitudinal and lateral directions. In addition, state estimation performance through temporal integration of partial observation is improved by applying a gated current unit(GRU) to solve the Partially Observable Markov Decision Process (POMDP) problem of front camera images. Finally, we validate the performance of the proposed technique by deploying the trained network from the simulation to a real driving environment.