Towards Verified Safety-critical Autonomous Driving Scenario with ADSML

Abstract

Modeling and verifying safety-critical scenarios of Autonomous Driving System (ADS) have increasingly attracted attention from academy and industry. The major challenge is lacking the domain-specific modeling language for ADS. To deal with this problem, we design and implement an Autonomous Driving Scenario Modeling Language (ADSML) based on the domain knowledge. The metamodel of ADSML describes the modeling elements and their relationships, which is used to capture the specific features of scenario. The concrete syntax of ADSML makes it easy to specify complex relationships among scenario elements, more important, we propose the contract module of ADSML to model the dynamic aspects of scenario. We use the semantics of Stochastic Hybrid Automata (SHA) to specify the dynamic behaviors in scenarios, which is seamlessly integrated with the model checker UPPAAL-SMC. With the help of the automatic model transformation, the ADSML models can be verified with UPPAAL-SMC to analyze the behaviors in scenarios. To demonstrate the feasibility, the scenario of lane change overtaking is modeled and some safety-critical properties are analyzed. The novelty of our approach is that it integrates the advantages of visual modeling and formal modeling. It helps the designers to model and verify the scenario models of autonomous driving systems.