Autonomous Drone Navigation with a Tiny Neural Network

Researchers developed a two-part system to enable a drone to fly autonomously indoors without GPS or expensive sensors. The 'teacher' pipeline uses computer vision techniques to detect and track a target, while the 'student' is a compact 1.7M parameter neural network that directly predicts velocity commands from camera images.

Details

The researchers tackled the problem of enabling autonomous drone flight indoors without relying on GPS, expensive sensors like LiDAR, or fiducial markers. They developed a two-part system: a 'teacher' pipeline that uses computer vision techniques to detect and track a target, and a 'student' neural network that learns to directly output velocity commands from raw camera images. The teacher pipeline uses YOLOv11 Nano for object detection and a U-Net 'Mask Splitter' to determine the front and back of the target, enabling consistent keypoint ordering for the IBVS control law. This teacher pipeline runs at 48 FPS. The student is a compact 1.7M parameter convolutional neural network trained on image-velocity pairs generated by the teacher in a digital twin simulator. This student network can run onboard the drone, enabling autonomous indoor flight without any GPS or markers. The researchers presented this work at ICCV 2025, one of the top computer vision conferences.