This paper presents the FalconWing, an ultra-light (150g) indoor fixed-wing UAV platform for vision-based autonomy. Indoor environments enable year-round, repeatable UAV experiments, but impose strict weight and maneuverability constraints on UAVs, driving the ultra-light FalconWing design. The FalconWing combines a lightweight hardware stack (a 137g airframe and a 9g camera) with off-board computing, and features a software stack featuring a realistic 3D Gaussian Splat (GSplat) simulator for vision-based controller development and evaluation. In a leader-follower case study, the best vision-based controller, trained using imitation learning and adding domain randomization to GSplat-rendered data, achieved 100% tracking success across three types of leader maneuvers in 30 trials, demonstrating robustness to leader appearance changes in simulation. In an autonomous landing case study, a vision-based controller trained purely in simulation was transferred to real hardware with zero-shot accuracy, achieving 80% success rate across 10 landing trials. FalconWing will release its hardware design, GSplat scene, and dynamic models to enable it to become an open-source flight kit for engineering students and research labs.
