Receding Horizon Weather Navigation Using Learned Navigation Policies

Navigating weather presents a significant challenge for pilots, demanding substantial training and experience. Prior research has explored using imitation learning to learn navigation policies that generate weather-aware flight routes based on off-line pilot demonstrations. However, the dynamic nature of weather and the evolving aircraft state make executing these learned policies difficult. This paper proposes an iterative method for generating and implementing learned policies over a receding horizon. By employing a receding horizon strategy, a new route is generated at each cycle, considering the latest environmental and aircraft conditions. Only a small segment of the route is executed before the process repeats until the destination is reached. This frequent updating of states enhances adaptability and resilience to weather variations and update delays, as demonstrated in simulations using NEXRAD data. Initial findings, reviewed and evaluated by our test pilots, suggest a promising avenue for this research.