Title: Testing Autonomous Driving Systems

Abstract:  Recent years have seen rapid progress in Autonomous Driving Systems (ADSs). To ensure the safety and reliability of these systems, extensive testing is required. However, direct testing on the road is incredibly expensive and unrealistic to cover all critical scenarios. A popular alternative is to evaluate an ADS’s performance in some well-designed challenging scenarios, a.k.a. scenario-based testing. Such test cases must possess several desirable properties (e.g., failure-inducing, realistic, etc.) to be useful. However, the search space of such test cases can be huge due to the temporal nature of traffic scenarios. In this talk, I will cover our recent efforts in efficiently generating testing scenarios: 1) AutoFuzz, a grammar-based, learning-guided black-box fuzzing technique to generate failure-inducing scenarios for ADSs; 2) FusED, an evolutionary and causality-based domain-specific grey-box fuzzing framework to generate failure-inducing scenarios for fusion component of ADSs; and 3) CTG, a Signal Temporal Logic (STL) guided conditional diffusion model that generates realistic and user-controllable scenarios for ADSs.

Bio: Baishakhi Ray is an Associate Professor in the Department of Computer Science at Columbia University, NY, USA. She has received the prestigious IEEE TCSE Rising star award and NSF CAREER award. Baishakhi’s research interest is in the intersection of Software Engineering and Machine Learning. Her research has been acknowledged by many Distinguished Paper awards and has also been published in CACM Research Highlights, and has been widely covered in trade media.