Deep reinforcement learning-based control algorithm for flight kinematics of insect-scale flyers

Abstract

An autonomous flight control algorithm based on deep reinforcement learning (DRL) is developed for insect-scale flyers with flexible wings in complex flow environments, addressing the challenges posed by highly unsteady and nonlinear aeroelastic dynamics. Unlike conventional model-based approaches, this study employs high-fidelity computational fluid–structural dynamics (CFD-CSD) simulations that fully resolve the governing equations of both the fluid and the flyer, providing physically consistent data for training the DRL agent. To mitigate the computational cost, a novel physics-guided data augmentation strategy is introduced, which synthetically expands the training dataset by replicating CFD-CSD data across diverse virtual flight scenarios while preserving the underlying physics. This approach enables the DRL agent to learn a robust control policy that generalizes across a broad range of aerodynamic conditions, demonstrating strong control performance even in complex and untrained flow environments. This work establishes a scalable framework for the autonomous control of flexible, bio-inspired flyers under realistic aerodynamic conditions, representing a significant step toward fully autonomous insect-scale flight.