Search-based Test-CASe Generation by Monitoring Responsibility Safety Rules

Abstract

The safety of Automated Vehicles (AV) as Cyber-Physical Systems (CPS) depends on the safety of their consisting modules (software and hardware) and their rigorous integration. Deep Learning is one of the dominant techniques used for perception, prediction and decision making in AVs. The accuracy of predictions and decision-making is highly dependant on the tests used for training their underlying deep-learning. In this work, we propose a method for screening and classifying simulation-based driving test data to be used for training and testing controllers. Our method is based on monitoring and falsification techniques, which lead to a systematic automated procedure for generating and selecting qualified test data. We used Responsibility Sensitive Safety (RSS) rules as our qualifier specifications to filter out the random tests that do not satisfy the RSS assumptions. Therefore, the remaining tests cover driving scenarios that the controlled vehicle does not respond safely to its environment. Our framework is distributed with the publicly available S-TALIRO and Sim-ATAV tools.

Too Afraid to Drive: Systematic Discovery of Semantic DoS Vulnerability in Autonomous Driving Planning under Physical-World Attacks

“…These methods assume differentiability of the test subject, which thus cannot be applied to the more industry-representative program-based BP targeted in this paper ( §II-A). Previous works also falsify safety properties on AD/RV software [10,28,[76][77][78][79][80]143]. They can not be directly applied since the problem scopes are different and the guidance is only limited to black-box guidance.…”

“…Recently, property-based testing have achieved great success in discovering safety violations in AD software [10,28,[76][77][78][79][80]. We follow the same general framework to develop a tool which can systematically discover DoS vulnerabilities in AD software.…”

“…For automated software vulnerability discovery in general, existing dynamic testing methods popularly obtain code-level feedback to guide the discovery process, which shows superior effectiveness over treating the software as a black-box [82][83][84][85][86]. In the general CPS testing domain, prior works have used quantitative feedback such as the robustness metrics to guide testing [28,79], but such a guidance still treats the code-level decision logic (i.e., decision-making code branches) as blackbox, which thus cannot provide guidance once the overall planning output stays unchanged. In our problem setting, it is desired if we can improve this with more fine-grained codelevel guidance such as the distance between the current inputs and the planning decision boundary at the code branch level.…”

“…As illustrated in Fig. 6, we fix the position of one static object right next to the left lane line and 15m in front of the AD vehicle, and at the same time, we arbitrarily move the other static object and measure the robustness distance metric used in [28,79,80] and BP vulnerability distance proposed by our paper. In Fig.…”

A Survey of Algorithms for Black-Box Safety Validation of Cyber-Physical Systems

“…Failures can include collisions with pedestrians (Koren & Kochenderfer, 2020) or other vehicles . Failures may also include violations of traffic laws (Kress-Gazit & Pappas, 2008) or other rule-sets, such as those designed to prevent at-fault collisions (Hekmatnejad et al, 2020).…”