Rare Collision Risk Estimation of Autonomous Vehicles with Multi-Agent Situation Awareness

TL;DR

This paper introduces a formal framework for estimating rare collision risks in autonomous vehicles using multi-agent situation awareness and advanced stochastic simulation techniques, enhancing safety assessment accuracy in complex environments.

Contribution

It presents a novel multi-agent risk estimation method combining GSHS modeling with IPS-FAS simulation to efficiently evaluate rare collision probabilities in autonomous driving scenarios.

Findings

Effective estimation of rare collision probabilities demonstrated in simulations.

Multi-agent situation awareness improves collision risk assessment accuracy.

The approach reduces computational costs compared to traditional simulation methods.

Abstract

This paper offers a formal framework for the rare collision risk estimation of autonomous vehicles (AVs) with multi-agent situation awareness, affected by different sources of noise in a complex dynamic environment. In our proposed setting, the situation awareness is considered for one of the ego vehicles by aggregating a range of diverse information gathered from other vehicles into a vector. We model AVs equipped with the situation awareness as general stochastic hybrid systems (GSHS) and assess the probability of collision in a lane-change scenario where two self-driving vehicles simultaneously intend to switch lanes into a shared one, while utilizing the time-to-collision measure for decision-making as required. Due to the substantial data requirements of simulation-based methods for the rare collision risk estimation, we leverage a multi-level importance splitting technique, known as interacting particle system-based estimation with fixed assignment splitting (IPS-FAS). This approach allows us to estimate the probability of a rare event by employing a group of interacting particles. Specifically, each particle embodies a system trajectory and engages with others through resampling and branching, focusing computational resources on trajectories with the highest probability of encountering the rare event. The effectiveness of our proposed approach is demonstrated through an extensive simulation of a lane-change scenario.