Tradeoff When Considering Deep Reinforcement Learning for Contingency Management in Advanced Air Mobility: Part II Risk Reduction to Populated Areas

With the introduction ofAdvanced Air Mobility (AAM) comes a heterogeneous set of vehicle capabilities and densities not previously seen in the airspace. It is likely that automation of these operations will be required to maintain the level of safety and sufficiently increase efficiency in this complex airspace. Artificial Intelligence (AI) can provide autonomous agent policies with real-time decision-making capabilities that can monitor the evolving risk of interrelated (or independent) hazards and, if necessary, execute appropriate control interventions. Building on prior work on contingency management, this work compares the tradeoffs of using Deep Reinforcement Learning (DRL) in addressing safety and efficiency of AAM operations. This paper extends previous research by incorporating population density as a dynamic factor in real-time risk assessment, refining a contingency management framework to reroute aircraft away from high-risk zones and achieving higher operational suitability. Through curriculum learning, agents learn to execute rerouting strategies under diverse operational conditions, including uncertainty in energy reserves, navigation hazards, and population density shifts. A comparison of these learning-based agents and a refined classical technique is presented in terms of their performance, verification difficulties, and development process.