Flapping-Wing UAV Autonomous Obstacle Avoidance
Bio-inspired monocular perception, image stabilization, reinforcement learning, and real-flight validation for a bird-like flapping-wing micro UAV.
This project studied autonomous obstacle avoidance for a bird-like flapping-wing micro aerial vehicle. Unlike multirotor UAVs, flapping-wing aircraft have severe constraints in size, payload, onboard computation, and image stability, while still needing fast perception, planning, and control during forward flight.
The platform had a total takeoff weight of about 250 g, a payload budget of about 20 g, and a flight speed of around 10 m/s. These constraints made common stereo or depth-camera solutions unsuitable, so the system used a lightweight monocular camera and a bio-inspired perception pipeline.
Key technical components:
- Bio-inspired monocular obstacle perception based on the LGMD mechanism, inspired by insect compound-eye neural responses to looming objects.
- IMU-assisted image stabilization to reduce flapping-induced pitch oscillation and improve monocular perception reliability.
- AirSim-based simulation training for constant-altitude obstacle avoidance, where the UAV avoided trees through roll commands while flying toward a target.
- Deep reinforcement learning policy training and deployment to real-world flight tests.
- Real-flight validation on the “Xinge” flapping-wing UAV, with 10 minutes of flight time and 14 autonomous-mode switches, all completing obstacle avoidance successfully.