Abstract:
Automated Parking System (APS) is a modern vehicle-equipped AI system that automates the process of parking vehicles. Nowadays, various companies (e.g., Tesla) have already deployed APSs on their latest released vehicles. Given the popularity of APSs, however, real-world APS misbehaviors (e.g., collision) continue to occur, calling for reliable techniques for the robustness testing of APSs. Existing works generally focus on safety testing of Autonomous Driving Systems (ADS) on public roads, which cannot comply with the unique characteristics of parking scenarios (e.g., vehicle behaviors and testing criteria). In light of this, we propose APSFUZZ, a novel simulation-based APS fuzzer to effectively detect bugs that result in misbehaviors (e.g., collision, stuck, pose error, etc.) of APS. Based on the systematic modeling of parking scenarios, APSFUZZ leverages parking-scenario-specific mutation strategies and a scheduling mechanism to ensure the effectiveness of fuzzing-based simulation testing. In the evaluation, we built the prototype of APSFUZZ based on the Carla simulator to identify the robustness flaws of Autoware.Universe (i.e., an open-source APS). Finally, APSFUZZ helped identify 74 buggy parking scenarios for Autoware. Universe, caused by 5 types of root causes. We have reported these 5 root causes to the developers, and till now 1 of them has been patched.