Propeller Optimization of Micro Flying Robots by Deep Reinforcement Learning

The propeller design optimization of a magnetically driven subcentimeter flying robot has been developed by using a deep reinforcement learning (DRL) scheme. The untethered rotating-wing robot is actuated by a single axis alternating magnetic field for diverse navigable flights. Advancements as compared to prior state-of-the-art works include: 1) an airfoil design optimization framework by deep reinforcement learning for tiny rotating-wing flying robots; 2) a very efficient exploration and identification algorithm with versatile geometric parameterizations; and 3) a 16% improvement in the lift-to-drag ratio for a 9.4 mm robot. As such, this work exemplifies a new class of machine learning-based design methodology for micro flying robots, which is also suitable for the design of a wide range of other micro actuators and robotic systems.