Perching Dynamics of Landing Gears in Drones over Asymmetrical Surfaces Using Deep Reinforcement Learning

Abstract

Perching of drones over asymmetrical surfaces leads to catastrophic damage especially if the breakdown goes undiscovered. By obtaining high pitch rates to take advantage of dynamic stall, a UAV with a variable sweep wing can conduct a perched landing on the ground. For natural and human-powered fliers, controlled gliding is one of the most energy-efficient ways of transportation. We show that without explicit knowledge of the underlying physics, gliding and landing methods with varied optimality criteria may be found via deep reinforcement learning. We use a two-dimensional model of a controlled elliptical body in conjunction with deep-reinforcement-learning (D-RL) to achieve gliding at a preset point with the least amount of energy expenditure or the quickest arrival time.To calculate success and failure acceleration differences and determine decision durations, an analytical model was constructed. Two distinct decision periods, corresponding to properly engaging the gripper and overloading the gripper's capabilities, have been demonstrated to be effective. The most relevant features within these two time periods, according to a machine learning feature selection method, are the quadrotor's maximum Z axis acceleration and the presence of near-zero readings. The gliding trajectories are smooth in both situations, while small/high frequency actuations differentiate the energy/time optimal techniques.We show that D-RL gliders may generalize their tactics to reach the goal destination from previously unknown starting points. D-model-free RL's nature and robustness suggest it could be a suitable framework for creating robotic devices that can exploit complex flow conditions.