Adaptive Methods for Fixed-Wing Aerial Vehicles

Autonomy in fixed-wing airplanes provide valuable services to modern life and have seen increasing use over the recent years. The ability to adapt to changing scenarios is critical for consistent performance of autonomous agents, and this ability requires unique solutions to overcome. This work presents several methods of improving adaptability of autonomous fixed-wing aircraft. We propose three aims, all applied on fixed-wing aircraft, as follows: 1. improving adaptability in primitive-based planners by a novel method of connecting maneuver primitives from the Maneuver Automaton, 2. augmenting a learned model-free policy with search to increase adaptability of the trained model-free policy, and 3. generate an "adaptive pilot" capable of adjusting to its observed trajectory through Domain Randomization, a reinforcement learning technique. The proposed work will be demonstrated as follows: 1. automatic generation of granular motion primitives that improve re-planning performance, 2. use of Monte Carlo Tree Search with a trained model-free policy to improve performance in no-fly zones and dynamic model changes, and 3. robustness to significant model parameter changes in the trained reinforcement learning model.