Collin Craig
Back to the wall

How it works

Precision Landing Lab

Teaching a drone to land itself on a rooftop pad, and to know when to abort and go around.

Open sourceView on GitHubContributions welcome

Stage
Early. My landing code flies in the simulator next.
Runs in
ArduPilot’s simulator
My part
The landing logic, by hand
License
Apache-2.0
Concept animation. A drone approaches a rooftop pad and looks for it with its downward camera.
The animation shows the plan, not flight data.

The job

A drone must find a rooftop pad and land on it. If it loses the pad, it must abort, climb and go around. It all runs in simulation.

The logic

The mission is a set of states with two paths: land, or abort and go around. This is the part I write.

  • Approach
  • Pad detected
  • Descend
  • Land
  • Target lost
  • Abort · climb
  • Go around

Two parts

The work is split so you can see exactly what is mine.

Part 1 · agent-built

The platform

  • Simulation tooling
  • Replay viewer
  • Run validator
  • Automated tests

Public. AI agents build it, and I direct and review the work.

Part 2 · my own code

Landing logic

  • Mission state machine
  • Frame chain
  • Estimator
  • Abort logic

Not public yet. I write it myself, with no AI-written code. It goes public with its first simulated flight.

What works today

  • A scripted mission in the simulator: takeoff to 5 m, hold, land (ArduPilot Copter 4.7.2).
  • A log exporter for the simulator's messages (MAVLink).
  • A replay viewer, built with three.js.
  • A run validator, written in Rust.
  • Automated tests (CI).

How it is tested

Random seeds are registered before the runs. Some tests are held out. Results will come with confidence intervals, and failures will be reported.

There are no landing-accuracy numbers yet. They go here after the first simulated flights with my landing code.

Next

Milestone 1: the first flight with my landing code, in the simulator.