May I Cut Into Your Lane?: A Policy Network to Learn Interactive Lane Change Behavior for Autonomous Driving

Abstract

In this paper, we propose a new lane change policy network which can learn interactive lane change behavior using reinforcement learning. The proposed policy network decides whether it will change lanes or not based on the behavior of the neighboring vehicles. First, the perception module obtains the relative location and speed of the neighboring vehicles along with the detected lane marking and encodes them into two dimensional synthetic image. Then, the convolutional neural network (CNN) is applied to extract the high-level interaction features from the input image and produce the lane change decision maximizing the expected accumulated reward. Then, the control module executes the appropriate driving maneuvers for carrying out the lane change decision. Note that the interactive driving policy learned in a goal (reward)-oriented manner produces the consistent driving behavior and generalizes well to diverse scenarios. Furthermore, separating the perception and control modules from the policy network allows the capacity of the whole network to be fully utilized to learn interactive lane change behavior. Our experiments show that our policy network trained with the deep Q network (DQN) framework yields safe and interactive driving behavior in a variety of scenarios.