Disentangling Perception-failure-induced Corner Cases by Counterfactual Reasoning

The reliability of the perception module of the autonomous driving system (ADS) faces tremendous and numerous challenges, although it has been verified following standards like ISO 26262. Therefore, finding safety-critical traffic scenarios is essential for testing autonomous driving perception. It is worth to note that, a corner case related to perception failure does not necessarily mean it is induced by perception failure. To address this notion, we propose to benefit scenario-based testing framework with counterfactual reasoning for disentangling the perception-failure-induced corner cases from the overall safety-critical scenarios. Taking the visual unawareness as the specific study case, we search for accidental testing scenarios with perception failures occurred simultaneously, by an LSTM-RNN controller from naturalistic driving data. Moreover, a combinational reward function is designed for accelerating this reinforcement discrete search process. A further verification in the simulated counterfactual world is then established to reason whether the perception failures induce the accidents or not. Experiments in the typical cases reveal the effectiveness of the proposed framework, which indirectly validates our notion as well.