FalconWing: An Ultra-Light Indoor Fixed-Wing UAV Platform for Vision-Based Autonomy

TL;DR

FalconWing delivers a 150 g indoor fixed-wing UAV with offboard computation and a GSplat photorealistic simulation to enable reproducible vision-based autonomy in GPS-denied environments. The authors demonstrate two challenging tasks: leader-follower tracking in simulation and zero-shot sim-to-real autonomous landing in indoor flight, achieving 100% success on unseen leader maneuvers in simulation and 80% landing success on hardware. The software stack combines photorealistic GSplat-based simulation, a nonlinear dynamics model, and an open-source package to support education and research, lowering barriers to hands-on airframe assembly and ROS-based vision pipelines. By providing an accessible, open-source flight kit, FalconWing enables iterative experimentation and benchmarking of vision-based control for ultra-light indoor fixed-wing platforms, with potential for broad adoption in education and labs.

Abstract

We introduce FalconWing, an ultra-light (150 g) indoor fixed-wing UAV platform for vision-based autonomy. Controlled indoor environment enables year-round repeatable UAV experiment but imposes strict weight and maneuverability limits on the UAV, motivating our ultra-light FalconWing design. FalconWing couples a lightweight hardware stack (137g airframe with a 9g camera) and offboard computation with a software stack featuring a photorealistic 3D Gaussian Splat (GSplat) simulator for developing and evaluating vision-based controllers. We validate FalconWing on two challenging vision-based aerial case studies. In the leader-follower case study, our best vision-based controller, trained via imitation learning on GSplat-rendered data augmented with domain randomization, achieves 100% tracking success across 3 types of leader maneuvers over 30 trials and shows robustness to leader's appearance shifts in simulation. In the autonomous landing case study, our vision-based controller trained purely in simulation transfers zero-shot to real hardware, achieving an 80% success rate over ten landing trials. We will release hardware designs, GSplat scenes, and dynamics models upon publication to make FalconWing an open-source flight kit for engineering students and research labs.

Figures (9)

• Figure 1: Left: Our ultra-light 150 g fixed-wing aircraft for indoor aerial research, equipped with a FPV camera and ROS-enabled autonomous control. Middle: Onboard view for leader-follower visual tracking using a digital camera. Right: Onboard view during autonomous landing using an analog camera.
• Figure 2: Architecture of FalconWing Hardware: a light 9 g FPV camera mounted on the fixed-wing plane streams images to the ground control station, where images are published to ROS. The controller reads published image plus buffered past controls, computes new flight control, and sends it via ROS to an Arduino. The Arduino writes these commands into the Spektrum NX-8 trainer port, closing the vision-based control loop over radio. The human pilot can instantly reclaim control at any time via a transmitter switch.
• Figure 3: FalconWing's simulation can render photorealistic images using Gaussian Splat from different poses. The top row shows 4 real world images in our flying arena, while the bottom row displays corresponding images rendered by GSplat at the same coordinates.
• Figure 4: iGSplat performance: top row shows one-shot pose estimation in analog-camera based simulation, while bottom row shows result in the digital-camera based simulation.
• Figure 5: Onboard camera views and trajectory plots for the leader-follower case study: our vision-based controller on the follower can closely track the leaders in three different leader maneuvers. The annotated red part on the onboard images indicating the mask detection result described in Section \ref{['sec:UMX-detection']}.