FalconWing: An Ultra-Light Fixed-Wing Platform for Indoor Aerial Applications

Yan Miao, Will Shen, Hang Cui, Sayan Mitra

Keywords: Sim-to-Real; Aerial Platform; Photo-realistic Simulation

Abstract: We introduce FalconWing, an ultra-light (150g) fixed-wing platform for indoor vision-based autonomy. Controlled indoor settings enable year-round, repeatable UAV experiments but impose strict mass and maneuverability limits, motivating a sensor-minimal design. FalconWing couples a lightweight hardware stack (137g UMX airframe, 9g analog FPV camera and offboard computation) with a world model composed of a photorealistic 3D Gaussian Splat (GSplat) simulation environment and a system-identified nonlinear dynamics. We validate FalconWing on two tasks without IMU or motion capture: leader-follower visual tracking and autonomous visual landing. Policies are trained via imitation entirely inside the world model, augmented with domain randomization over appearance and geometry. In simulation, our best learned policy attains 100% success across three unseen leader maneuvers and is robust to appearance shifts. For landing, a policy trained purely in the world model transfers zero-shot to hardware, achieving an 80% success rate over ten indoor trials. We will release hardware designs, GSplat scenes, dynamics parameters, and ROS workflows, positioning FalconWing as an open-source benchmark and educational kit for world-model-driven vision-based fixed-wing autonomy.

Submission Number: 16