Closed-Loop ACAS Xu Neural Network Verification

Benchmark Proposal: Neural Network Control Systems (NNCS) play critical roles in autonomy. However, verifying their correctness is a substantial challenge. In this paper, we consider the neural network compression of ACAS Xu, a popular benchmark usually considered for open-loop neural network verification. ACAS Xu is an air-to-air collision avoidance system for unmanned aircraft issuing horizontal turn advisories to avoid collision with an intruder aircraft. We propose specific properties and different system assumptions to use this system as a closed-loop NNCS benchmark. We present experimental results for our properties based on randomly generated test cases and provide simulation code.