Modeling and Analysis of Collision Avoidance with ACAS Xu in Verse for the GUAM Lift + Cruise Model

As advanced air mobility technology matures and becomes widely adopted, urban environments are expected to experience a high density of air traffic. To improve operational safety in dense traffic scenarios, collision avoidance protocols can be implemented in these air vehicles to generate high-level command advisories to avoid nearby vehicles or obstacles. Toward this end, some collision avoidance controllers have been implemented as data-driven modules, such as neural networks, to replace advisory lookup tables for storage and computational efficiency gains. However, small perturbations to the input of the neural network can potentially lead to different advisory outputs. Hence, neural network controllers must be verified to produce safe command advisories, which are then propagated through the vehicle dynamics for closed-loop verification purposes. In this work, we propose a framework for modeling and simulating a Lift+Cruise vehicle guided by collision avoidance advisories. The collision avoidance advisories are generated by a decision tree that approximates the existing ACAS Xu neural network system. The simulation is implemented in the Verse library to propagate the collision avoidance commands through the nonlinear Lift+Cruise vehicle dynamics in a closed-loop manner. This approach is used to generate simulations from a variety of initial vehicle conditions to find scenarios where ACAS Xu advisories can potentially lead to near mid-air collisions.