Safe Autonomy: Continuous Testing and Controller Compensation for Unreliable Perception

Abstract

This thesis addresses two key challenges in designing safe and reliable Autonomous Vehicles, evaluating the performance of vehicle controllers across different scenarios and designing safe vehicle controllers that can manage unreliable perception. Testing is essential for AVs as it helps identify potential flaws in controllers that could lead to unsafe actions, and it assists designers in debugging. However, testing is a complex process that involves integrating resources like simulators and various modules, including controllers and perception systems, which is often a cumbersome task. To simplify this, we propose a continuous testing pipeline that automates the evaluation of controllers in diverse simulation environments and provides developers with feedback to continuously improve their controllers. Our continuous testing pipeline has supported over 100 student developers from various universities, enabling them to test their controllers automatically. Designing controllers with learning-based perception is also crucial, given the popularity of learning-based perception for its scalability. Nevertheless, many existing controller synthesis algorithms assume perfect perception, overlooking the fact that perception can be unreliable and that perception errors can lead to unsafe control actions. To address this, we introduce a two-step approach, an offline phase that identifies perception uncertainties using a preimage perception contract, followed by the real-time implementation of a risk heuristic to ensure safe control. Our simulations, conducted in various adaptive cruise control scenarios across different tracks and weather, have demonstrated the effectiveness of this strategy, resulting in a 73% reduction in safety violations such as collisions and lane departures caused by unreliable perception.

Type
Publication
Master Thesis