Northstar

Inspiration

Drones are everywhere, but safety is always the afterthought. We kept seeing autonomous systems that were impressive until something went wrong, and when it did, there was nothing stopping a bad command from reaching the vehicle. We wanted to flip that. Northstar started with one question: what if safety was the first thing the drone checked, not the last?

What it does

Northstar is an autonomous drone system where every command from every agent passes through a Safety Gate before the drone acts on it. The gate checks geofencing, altitude, LIDAR obstacle readings, and battery state in real time. If a command is unsafe, it gets blocked or modified before it ever reaches the vehicle. A live Cesium 3D visualization shows exactly what's happening so you can see the safety decisions as they're made.

How we built it

We built a full drone simulator running on the DGX Spark that speaks the same MAVLink-compatible interface a real PX4 drone would use. We had a very strong idea to build a drone, but due to hardware difficulties, it made it impossible for us to get the drone to function. Multiple AI agents handle different jobs like mission planning, person following, battery management, and orchestration, and they all talk to the drone through the same Safety Gate. The sim runs at 20Hz with forward and downward LIDAR and a battery model baked in.

Challenges we ran into

Getting the sim, safety gate, and agent stack to stay in sync without a real flight controller was harder than expected. Cesium tile loading on conference WiFi was a nightmare, so we ended up caching tiles locally. Designing the MAVLink interface to be hardware portable without using real pymavlink serialization required some careful thinking up front.

Accomplishments that we're proud of

The Safety Gate, being the single chokepoint for every agent command, is the thing we're most proud of. Obstacle avoidance and policy enforcement are literally the same code path.

What we learned

We learned that doing the simulation first is the right call. We initially started off with building a physical drone, which was going to be powered by an FC(flight controller) and a Raspberry Pi 5. Due to a very small misunderstanding in our hardware, we weren't able to connect the receiver to the remote controller. This was quite frustrating because we had set up the rest of the hardware and wiring with no problem. We then had to rush to build our simulation, which was definitely challenging.