Highway Lane change under uncertainty with Deep Reinforcement Learning based motion planner

Abstract

Motion Planning is a fundamental component of a mobile robot to reach its goal safely avoiding collision. For a self-driving car on a highway, the presence of non-communicating vehicles, specially those whose intent is unknown, creates a lot of uncertainty for the motion planner in generating a safe trajectory. State-of-the-art planning methods do not work well in case of adversary driving scenarios, where the other vehicles may make mistakes or have a competing or malicious intent. We use reinforcement learning framework to improve safety under those scenarios. In most recent deep reinforcement learning applications, there is a neural network that maps an input state to an optimal policy over actions. However, learning a policy over such original or primitive actions is very slow and inefficient and is therefore not suitable for many robotics tasks. On the other hand, the knowledge already learned in classical planning methods should be inherited and reused. In this thesis, in order to take advantage of reinforcement learning good at exploring the action space for optimal solution and classical planning skill models good at handling most driving scenarios, we propose to learn a policy over an action space of primitive actions augmented with classical planning methods. By doing so, we show that our agent outperforms the primitive-action reinforcement learning agent and the classical planning methods in terms of collision rate