Data-Driven Control of Insect Flapping Flight via Deep Reinforcement Learning.

Area of Science:

• Robotics and Biomechanics
• Computational Fluid Dynamics
• Bio-inspired Engineering

Background:

• Simulating insect flight is complex due to intricate wing movements and aerodynamic forces.
• Existing models often struggle to capture the full dynamics of insect flight.
• Miniature insect flight presents unique challenges in modeling kinematics and aerodynamics.

Purpose of the Study:

• To develop a bidirectional simulation framework for miniature insect flight.
• To couple wing kinematics with aerodynamic forces for realistic simulation.
• To enable autonomous flight control and task performance in simulated insects.

Main Methods:

• Parameterizing natural wingbeat cycles from real-world data for insect kinematics.
• Utilizing an improved semi-empirical model for computing aerodynamic forces, including unsteady components.
• Employing deep reinforcement learning for adaptive control of flapping strategies.
• Integrating a controller for autonomous wing motion regulation and task execution.

Main Results:

• The framework successfully models the bidirectional coupling of kinematics and aerodynamics.
• Deep reinforcement learning enabled adaptive flapping strategies for dynamic flight states.
• The simulated insect demonstrated autonomous control for tasks like obstacle avoidance.
• Generated flight is physically plausible and autonomous across various scenarios.

Conclusions:

• The developed framework effectively simulates realistic and autonomous miniature insect flight.
• This approach advances the field of bio-inspired robotics and flight simulation.
• The integration of deep reinforcement learning offers a powerful tool for adaptive control in simulated biological systems.