DominIQ

Inspiration

We set out to solve a hard physical AI problem with real-world impact in the defense and maritime surveillance sectors. The challenge: autonomously detect, locate, and continuously track a moving vessel across a large Arctic environment using only drone-mounted cameras and onboard sensors. No GPS on the target, no pre-known position, no human operator in the loop.

What it does

DominIQ autonomously detects, locates, and tracks a moving vessel across a 6.5km Arctic environment using coordinated drones and AI. It fuses YOLO detections with real-time drone telemetry to compute the boat's GPS coordinates purely from camera geometry, with no human in the loop.

How we built it

We built a multi-asset MAVLink controller that flies a quadcopter and fixed-wing plane in coordinated patrol orbits while running a custom-trained YOLO model on their live camera feeds. On top of that we layered a voice interface using Whisper, GPT-4o, and ElevenLabs so operators can ask questions about the mission in plain English and get spoken answers back.

Challenges we ran into

Figuring out how to position and deploy the drones for optimal performance was somewhat difficult. It was unclear at first, but after experimenting over the weekend, we developed a reasonable arrangement.

Accomplishments that we're proud of

We're proud that the full pipeline works end-to-end with zero human guidance: the drones take off, find a randomly spawned moving vessel somewhere in the site, and lock onto it using only what the camera sees. The voice assistant accurately answers live mission questions because it reads real telemetry logs, not canned responses.

What we learned

Physical AI is a completely different beast from pure software AI because every sensor has noise, every motor has lag, and the math that looks clean on paper breaks down the moment the drone tilts 5 degrees. We learned that multi-agent coordination over MAVLink requires extremely careful message sequencing since shared state and concurrent reads will silently corrupt your flight loop in ways that are almost impossible to debug.

What's next for DominIQ

The next step is deploying on real hardware with actual ArduPilot autopilots and physical gimbaled cameras, which our architecture already supports since we built against standard MAVLink throughout. We also want to add triangulation from multiple drones simultaneously so the GPS estimate improves as more cameras observe the target from different angles.

Built With

• html5
• python
• typescript