Safe and Goal-Based Highway Maneuver Planning with Reinforcement Learning

Abstract

As autonomous driving moves closer to a real-world application, more and more attention is being paid to the motion planning part of the system. To handle vastness of possible road scenarios, negotiate with other road users and generate an intelligent control strategy in a constantly changing environment, data-driven techniques and artificial intelligence methods seem to be the approach of choice. In this paper, we present reinforcement learning (RL) agent which is embedded in a deterministic, safety envelope. The agent is responsible for generating high-level maneuvers, such as a lane following or a lane change. The primary goal of the agent is to reach a given lane in a given distance, while traveling on a highway. The selected maneuver is then executed with use of deterministic methods utilizing concept of Responsible-Sensitive Safety (RSS) framework, which formalizes safety constrains in a form of mathematical model. The proposed solution has been evaluated in two environments: one in which the agent receives a predefined reward for getting to a correct lane and second, in which it is rewarded for doing this in a time-optimal manner. We have evaluated the proposed solution against an another RL-based agent, which is steering vehicle by low-level control signals, such as acceleration and steering angle.