Towards Automated Driving Violation Cause Analysis in Scenario-Based Testing for Autonomous Driving Systems

TL;DR

This work introduces DVCA, a tool for automated driving violation cause analysis in scenario-based ADS testing. By deploying idealized substitutes within a simulator and leveraging counterfactual causality, it attributes system-level violations to specific ADS components and messages. Evaluations on real-bug and injected-fault benchmarks show 100% component-level attribution and >98% message-level attribution, with substantial reductions in debugging scope and practical efficiency gains. The study demonstrates the value of simulation-driven counterfactual analyses for rapid, targeted debugging in autonomous driving systems.

Abstract

The rapid advancement of Autonomous Vehicles (AVs), exemplified by companies like Waymo and Cruise offering 24/7 paid taxi services, highlights the paramount importance of ensuring AVs' compliance with various policies, such as safety regulations, traffic rules, and mission directives. Despite significant progress in the development of Autonomous Driving System (ADS) testing tools, there has been a notable absence of research on attributing the causes of driving violations. Counterfactual causality analysis has emerged as a promising approach for identifying the root cause of program failures. While it has demonstrated effectiveness in pinpointing error-inducing inputs, its direct application to the AV context to determine which computation result, generated by which component, serves as the root cause poses a considerable challenge. A key obstacle lies in our inability to straightforwardly eliminate the influence of a specific internal message to establish the causal relationship between the output of each component and a system-level driving violation. In this work, we propose a novel driving violation cause analysis (DVCA) tool. We design idealized component substitutes to enable counterfactual analysis of ADS components by leveraging the unique opportunity provided by the simulation. We evaluate our tool on a benchmark with real bugs and injected faults. The results show that our tool can achieve perfect component-level attribution accuracy (100%) and almost (>98%) perfect message-level accuracy. Our tool can reduce the debugging scope from hundreds of complicated interdependent messages to one single computation result generated by one component.

Figures (6)

• Figure 1: Overview of representative ADS systems and the interactions between its internal components.
• Figure 2: Illustration of collision between vehicle controlled by the AD software and the pedestrian. (a). shows the moving trajectories of both the vehicle and the pedestrian in the accident. (b)-(e) shows the autonomous vehicle's classification and prediction results of the pedestrian and its corresponding decision. Figures are generated based on the description in NTSB's investigation report ntsb_report.
• Figure 3: Illustration of working pipeline of our DVCA tool when analyzing a prediction-induced fault under a collision scenario.
• Figure 4: Illustration of the idealized substitute. Blue arrows are used to highlight idealized substitutes within the data flow. To provide a concrete example of this, we present a specific instance of the data flow, illustrating the information transmitted in each idealized substitute.
• Figure 5: Execution difference between two traces collected from the same testing scenario. Each node represents the driving distance and relative time when the planning component is executed.