Deep Reinforcement Learning Based Collision Avoidance of Automated Driving Agent

Automated driving has become a very promising research direction with many successful deployments and the potential to reduce car accidents caused by human error. Automated driving requires automated path planning and tracking with the ability to avoid collisions as its fundamental requirement. Thus, plenty of research has been performed to achieve safe and time efficient path planning and to develop reliable collision avoidance algorithms. This paper uses a data-driven approach to solve the abovementioned fundamental requirement. Consequently, the aim of this paper is to develop Deep Reinforcement Learning (DRL) training pipelines which train end-to-end automated driving agents by utilizing raw sensor data. The raw sensor data is obtained from the Carla autonomous vehicle simulation environment here. The proposed automated driving agent learns how to follow a pre-defined path with reasonable speed automatically. First, the A* path searching algorithm is applied to generate an optimal path from origin to destination for the agent represented by waypoints. Further, preview path tracking errors, steering control and distance to destination are introduced to build the reward function. In addition, raw data from multiple sensors is processed separately and concatenated together to help the proposed agent get a comprehensive understanding of its environment. Two traffic scenarios including traffic rule free urban road and road segment with two intersections, traffic light and stop sign are used to evaluate the performance of the proposed automated driving agent. The performance of proposed Deep Q-Learning (DQN) agent is evaluated in multiple aspects. Compared to traditional mid-to-mid DRL agent with explicit decomposition of high-level behavior decision and low-level control, the proposed DRL agents are expected to have better performance and smaller size since all processing steps are optimized simultaneously. Moreover, the pre-calculated A* path provides a good reference point for subsequent DRL training.