Integrating Throttle into a Reinforcement Learning Controller for a Perched Landing of a Variable Sweep Wing UAV

This paper details the training, and simulation and experimental results for controlling a perching manoeuvre of a variable sweep uncrewed air vehicle (UAV) using reinforcement learning (RL). This builds on previous work by the authors by allowing the RL controller to determine the throttle cut-off point during the initial phase of the manoeuvre. This additional control allows the system to show significantly improved performance over the baseline models. In simulation the addition of throttle showed a 5x improvement in performance under a 8 m/s headwind. Similar improvements were seen in challenging real-world conditions with significant headwinds. Further testing of the trained agents under varying real-world flight conditions would help confirm initial results. A further degree of throttle control may also improve performance but will require a significantly improved numerical model.