End-to-end Autonomous Driving in Heterogeneous Traffic Scenario Using Deep Reinforcement Learning

In this paper, we propose an end-to-end autonomous driving architecture for safe maneuvering in heterogeneous traffic using a reinforcement learning (RL) algorithm. Using the proposed architecture we develop an RL agent that can make driving decisions directly from the sensor data. We formulate the autonomous driving problem as a Markov Decision Process and propose different architectures using Deep $Q -$Networks for two types of sensor data - top view images of the autonomous vehicle (AV) and its surrounding vehicles and information on relative position and velocities of the surrounding vehicles w.r.t the AV. We consider a highway scenario and analyze the performance of the RL agent using the proposed architectures using the highway-env simulator. We compare the driving performance of the AV for both sensor types and discuss their efficacy under varying traffic densities.