Towards End-to-End Escape in Urban Autonomous Driving Using Reinforcement Learning

Mustafa Sakhai, Maciej Wielgosz

Abstract: This paper addresses the challenging task of developing an autonomous escape protocol in self-driving cars. An autonomous chase protocol was created to test self-driving cars’ escape scenarios. Escape should be understood as getting to a certain point in the shortest possible time with the addition of considering changing environmental conditions. First, an autonomous vehicle capable of driving autonomously from point A to B was developed from scratch. We used a dedicated curriculum learning agenda which allowed the proposed model to perform all fundamental road maneuvers. We developed a discrete action space and a single RGB camera throughout the experiments. Based on the performed experiments, reward functions were proposed, which enabled practical training of the agent to make the right action at the right time. Furthermore, in the subsequent experiments, we selected the reward function and model that produced the best result, guaranteeing that the chasing car was within 25 meters of a runaway car for 63% of the episode duration. To the best of our knowledge, this work is the first one that addressed the task of the chase in urban driving using the Reinforcement Learning approach.