Improvements in learning to control perched landings

Abstract Reinforcement learning has previously been applied to the problem of controlling a perched landing manoeuvre for custom sweep-wing aircraft. Previous work showed that use domain randomisation train with atmospheric disturbances improved real-world performance controllers, leading increased reward. This paper builds on previous project, investigating enhancements and modifications process further improve performance, reduce final state error. These changes include modifying observation by adding information about airspeed standard aircraft vector, employing simulator, optimising underlying RL algorithm network structure, changing continuous action space. Simulated investigations identified hyperparameter optimisation as achieving most significant increase in reward performance. Several test cases were explored identify best combination enhancements. Flight testing was performed, comparing baseline model against some performing from simulation. Generally, performed better than simulation also real world. However, flight tests limitations current numerical model. For models, chosen policy performs well yet stalls prematurely reality, known reality gap.