ANTI-CARLA: An Adversarial Testing Framework for Autonomous Vehicles in CARLA

TL;DR

ANTI-CARLA is an automated adversarial testing framework for CARLA that identifies failure scenarios in autonomous driving systems by simulating challenging weather and sensor fault conditions.

Contribution

It introduces a comprehensive adversarial testing framework for CARLA, enabling systematic discovery of failure cases in autonomous driving pipelines.

Findings

Successfully identified failure scenarios despite high system accuracy

Automated search efficiently finds adversarial conditions

Extends CARLA with adversarial testing capabilities

Abstract

Despite recent advances in autonomous driving systems, accidents such as the fatal Uber crash in 2018 show these systems are still susceptible to edge cases. Such systems must be thoroughly tested and validated before being deployed in the real world to avoid such events. Testing in open-world scenarios can be difficult, time-consuming, and expensive. These challenges can be addressed by using driving simulators such as CARLA instead. A key part of such tests is adversarial testing, in which the goal is to find scenarios that lead to failures of the given system. While several independent efforts in testing have been made, a well-established testing framework that enables adversarial testing has yet to be made available for CARLA. We therefore propose ANTI-CARLA, an automated testing framework in CARLA for simulating adversarial weather conditions (e.g., heavy rain) and sensor faults (e.g., camera occlusion) that fail the system. The operating conditions in which a given system should be tested are specified in a scenario description language. The framework offers an efficient search mechanism that searches for adversarial operating conditions that will fail the tested system. In this way, ANTI-CARLA extends the CARLA simulator with the capability of performing adversarial testing on any given driving pipeline. We use ANTI-CARLA to test the driving pipeline trained with Learning By Cheating (LBC) approach. The simulation results demonstrate that ANTI-CARLA can effectively and automatically find a range of failure cases despite LBC reaching an accuracy of 100% in the CARLA benchmark.