Perceptual Risk-Aware Adaptive Responsibility Sensitive Safety for Autonomous Driving

Abstract

The Responsibility-Sensitive Safety (RSS) model is a state-of-the-art parametrizable approach to facilitating safety planning and control, which has been widely used in autonomous driving systems. However, the current RSS model neither considers perceptual risks, nor can adaptively adjust its parameter settings according to different scenarios. These limitations may lead to unsafe or inefficient behavior of the autonomous vehicles. Therefore, this paper proposes a novel perceptual risk-aware adaptive RSS approach, which trains the interpretable perceptual risk assessment model to evaluate the risk level of different scenarios and provides interpretable reasons for reference, then adaptively selects the corresponding parameters in the RSS model for safety monitoring according to the obtained perceptual risk level. This new risk-aware adaptive approach significantly reduces safety margins and increases traffic density, while maintaining risk limits. Our experiments illustrate that our approach can well balance the safety and practicality of autonomous driving systems for complex scenarios.