Formal Estimation of Collision Risks for Autonomous Vehicles: A Compositional Data-Driven Approach

Abstract

In this work, we propose a compositional data-driven approach for the formal estimation of collision risks for autonomous vehicles (AVs) with black-box dynamics acting in a stochastic multi-agent framework. The proposed approach is based on the construction of sub-barrier certificates for each stochastic agent via a set of data collected from its trajectories while providing a-priori guaranteed confidence on the datadriven estimation. In our proposed setting, we first cast the original collision risk problem for each agent as a robust optimization program (ROP). Solving the acquired ROP is not tractable due to an unknown model that appears in one of its constraints. To tackle this difficulty, we collect finite numbers of data from trajectories of each agent and provide a scenario optimization program (SOP) corresponding to the original ROP. We then establish a probabilistic bridge between the optimal value of SOP and that of ROP, and accordingly, we formally construct the sub-barrier certificate for each unknown agent based on the number of data and a required level of confidence. We then propose a compositional technique based on small-gain reasoning to quantify the collision risk for multi-agent AVs with some desirable confidence based on sub-barrier certificates of individual agents constructed from data. For the case that the proposed compositionality conditions are not satisfied, we provide a relaxed version of compositional results without requiring any compositionality conditions but at the cost of providing a potentially conservative collision risk. Eventually, we develop our approaches for non-stochastic multi-agent AVs. We demonstrate the effectiveness of our proposed results by applying them to a vehicle platooning consisting of 100 vehicles with 1 leader and 99 followers. We formally estimate the collision risk for the whole network by collecting sampled data from trajectories of each agent.