Application of an autonomous multi-agent system using Proximal Policy Optimisation for tactical deconfliction within the urban airspace

The present paper formalises the development of a Multi-agent Reinforcement Learning (MARL) solver for U-space Service Providers (USSPs) supporting the tactical conflict resolution and exhibited in the Air Mobility Urban - Large Experimental Demonstration (AMU-LED) project. It relies on an Advantage Actor Critic (A2C) model with a Proximal Policy Optimisation (PPO) learning baseline. The application of the autonomous system is demonstrated under a synthetic (with live and virtual) air/unmanned traffic management (ATM/UTM) environment. The Unmanned Aircraft Systems (UASs) are flying in cruise phase at low altitudes, whose respective flight plan generates intersections for enforcing a high collision frequency. The study adopts a step-wise complexity approach of scenarios that confront two agents' observation methods and showcases a practical case of tactical conflict resolution. The experiments show encouraging deconfliction performance with promising prospects for seeing a such solver deployed.