Single-person R&D focused on autonomous systems for UAVs that run in real time on affordable hardware.
UAV detection model
Preview of the model running on real-world UAV chasing video. It runs at 30-45FPS on the $150 hardware.
Drone Avoidance Module
Automatically maneuvers a UAV away from a drone approaching from behind. When a rear-facing camera detects the drone with high confidence, the module steers the UAV away from it. The video shows the system being tested in simulation.
The estimated price is $250-300, including a single-board computer, daylight camera, and all software. It may be lower when buying electronics in bulk.
RL Control Policy
A pure reinforcement learning model that controls an interceptor drone for more generalization and versatility. It outputs direct RC commands and does not rely on synchronized frame-IMU data, which avoids expensive hardware.
5GHz WiFi Broadcast HAT
An open-source 5GHz WiFi Broadcast HAT for FPV video transmission and long-range communications.
- Chipset: RTL8812EU in a BL-M8812EU2 module
- Frequencies: 5.15 to 5.85 GHz
The KiCad design files, driver information, and safety notes are available in the GitHub repository.
Tools
Here are some Open Source tools originally built for internal use. Feel free to check them out:
-
YOLO Auto Annotation creates YOLO dataset from specified images using Moondream AI model to find bounding boxes for your classes. It reduces time required to go through images and annotate them yourself.
-
YOLO Dataset Visualizer allows to scroll through dataset and see bounding boxes on images. You can remove image from dataset if bounding boxes are incorrect. It helps remove occasional mistakes after Auto Annotation.