A verification framework for behavioral safety of self-driving cars

While self-driving cars have already been widely investigated and achieved spectacular progress, a major obstacle in applications is the great difficulty in providing formal guarantees about their behaviors. Since the environment of the self-driving is usually not known beforehand and highly uncertain, classical verification approaches cannot be applied to guarantee safety. To cope with any traffic situation, a novel online verification framework is presented for verifying behavioral safety of self-driving cars. The framework is based on the proposed five safety considerations: new longitudinal and lateral safe distances, lane changes, overtaking and how to face new traffic participants. Different from the previous verification considerations, this verification framework allows actual behaviors of self-driving cars to be temporarily inconsistent with the popular strict safe distance. As long as the self-driving car respects the minimum safe distance calculated by our technique and executes improvement behaviors to restore the safe distance, it is still believed that the predictive behavior is safe. The framework can easily be integrated to existing self-driving systems and evaluate different indicators involving the steering angle, acceleration and braking. The benefits of the framework in different urban scenarios of the CARLA simulator and real traffic data provided by the NGSIM project are demonstrated. Results show that the technology can successfully detect unsafe behaviors and provide effective measures to avoid potential collisions.