Runtime Assurance of Learning-Based Lane Changing Control for Autonomous Driving Vehicles

Abstract

Learning techniques such as deep reinforcement learning have been increasingly used in the controller design for autonomous vehicles, e.g., lane changing controllers. Although the use of learning techniques can remarkably increase the level of autonomy, their presence poses a great challenge for safety assurance due to the black-box and data-driven nature of learning techniques. In this work, we study the problem of how to safeguard the learning-based lane changing controller for collision avoidance. Our solution leverages on the runtime assurance framework, in particular the Simplex architecture to bound the behavior of the learning-based controller and to provide safety guarantees. The basic idea is to encompass the learning-based controller with a safety-by-construction controller and a decision module, which monitors the output of the learning-based controller at runtime and implements a switching logic between these two controllers according to the changing environment. We present the detailed design of the decision module and formally prove its correctness for collision avoidance. We also carry out a comprehensive experimental evaluation in a set of realistic highway scenarios using the SUMO simulator. The simulation results show that our proposed solution can not only provide safety guarantee for the learning-based lane changing controller, but also maintain a considerable level of efficiency in different volumes of traffic flow.