ADR Platform

ADR Platform is experimental edge software for autonomous drone racing. It combines Hailo-accelerated gate detection, fisheye-camera pose estimation, browser-viewable video, UDP telemetry, and MAVLink mission control on a Raspberry Pi 5.

Caution

This is research software, not a safety-certified flight system. The mission code can send movement and landing commands to an aircraft. Develop against simulation first, test without propellers, keep a manual failsafe available, and follow local aviation and radio-control rules.

System overview

V4L2 camera or video
        |
        v
GStreamer + OpenCV --> Hailo-10H YOLO --> gate pose + web stream
                                             |
                                      UDP 127.0.0.1:5050
                                             |
                                             v
                                  navigation mission controller
                                             |
                                  MAVLink local-NED commands

The vision pipeline detects up to three gates, refines their image corners, estimates 3D pose with solvePnP, and publishes rows in this format:

[distance, forward, right, down, roll, pitch, yaw]

Positions are in metres, angles are in degrees, and 999.0 marks a missing detection.

Repository layout

• assets/ — model documentation and camera calibrations
• vision/ — capture, inference, pose estimation, telemetry, and Flask streams
• navigation/ — current modular MAVLink controller and mission framework
• scripts/ — supported entry points and recording helper
• debug/ — Hailo and UDP telemetry diagnostics
• Top-level navigation*.py files — legacy prototypes retained for reference

Hardware and software

The current configuration targets:

• Raspberry Pi 5 running 64-bit Raspberry Pi OS
• Raspberry Pi AI HAT+ 2 with a Hailo-10H accelerator
• A V4L2 camera capable of 1280×720 MJPEG at 60 FPS
• A MAVLink-compatible autopilot or simulator
• Python 3.10 or newer

Installation

Install the Hailo-10H runtime using the current Raspberry Pi AI software instructions:

sudo apt update
sudo apt install dkms hailo-h10-all
sudo reboot

After rebooting, verify the accelerator:

hailortcli fw-control identify

Install the remaining system dependencies:

sudo apt install \
  python3-venv \
  python3-gi \
  gir1.2-gstreamer-1.0 \
  gstreamer1.0-tools \
  gstreamer1.0-plugins-base \
  gstreamer1.0-plugins-good \
  gstreamer1.0-plugins-bad \
  gstreamer1.0-plugins-ugly

Clone the project and create a virtual environment that can access the system-installed Hailo and GStreamer Python bindings:

git clone https://github.com/ctrl-alt-delete101/ADR-platform.git
cd ADR-platform
python3 -m venv --system-site-packages .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

Configuration

Review these values before running anything:

1. In scripts/run_vision.py, set the camera device, resolution, and frame rate in CAMERA_CONFIGS.
2. Supply a Hailo-10H model at assets/best.hef. This local model is ignored by Git and is not distributed with the repository.
3. In vision/opencv_processing.py, set USE_VIDEO_FILE, VIDEO_FILE_PATH, and FLIP_CAMERA. Live-camera mode is the default.
4. Confirm that the selected calibration YAML matches the physical camera.
5. In navigation/navigation.py, verify MAVLINK_CONN, camera offsets, controller gains, speed limits, tolerances, and timeouts for your vehicle.

The included calibration files are hardware-specific and should not be assumed to fit another camera, lens, resolution, or mounting arrangement.

Running the vision pipeline

Start vision and telemetry:

python -m scripts.run_vision

Open http://localhost:5000 on the Raspberry Pi. The Flask development server listens on all interfaces and has no authentication or TLS. Do not expose it directly to the public internet.

To inspect telemetry without connecting an autopilot:

python -m debug.telemetry_reader

To record camera input:

./scripts/record.sh 0

Recordings are written to captures/, which is ignored by Git.

Running a mission

Only proceed after validating the vision output, coordinate frames, calibration, offsets, and MAVLink connection in a simulator:

python -m scripts.multi_stage_gate_mission
python -m scripts.single_gate_mission

The mission approaches each detected gate at staged distances and then commands a pass-through target. The single-gate mission instead takes 0.25 m receding-horizon steps, replans after every step, commits to the pass at 1 m, and lands after crossing. Treat the default controller values as development examples, not safe settings for an arbitrary vehicle.

To isolate camera-offset effects during multi-stage simulator testing, set the three corrections to zero from the command line:

python -m scripts.multi_stage_gate_mission \
  --camera-right-offset-m 0 \
  --camera-down-offset-m 0 \
  --camera-yaw-offset-deg 0

Model and calibration provenance

The compiled model is intentionally excluded from the repository. Users must supply a model they are licensed to use. The calibration files need separate provenance and redistribution review; see assets/README.md. Assets are not automatically covered by the source-code license.

Development

Run the hardware-independent checks:

python -m unittest discover -s tests -v
python -m compileall -q debug navigation scripts vision
bash -n scripts/record.sh

Contributions are welcome; see CONTRIBUTING.md. Report security issues privately as described in SECURITY.md.

License

Unless otherwise noted, the source code is available under the MIT License. Third-party dependencies and repository assets retain their own licensing terms.