Autonomous Drone Racer

The spark

The AI Grand Prix is a real autonomous drone racing league: standardized drones, unknown tracks, and a hard rule that the aircraft flies itself with no connectivity. I entered because GPS-denied navigation is the purest version of a problem I've worked on my whole career: how do you know where you are, and where to go, when the infrastructure you'd normally rely on is gone?

The problem

The AI Grand Prix racing league requires fully autonomous drones to thread race gates at speed with no GPS, no absolute position, and no human input: one camera and an IMU.

Who it affects

A competition entry, so the affected party is me and the leaderboard. The vision round forbids absolute position outright: monocular camera, IMU, and motor telemetry only.

Previous workflow

Most approaches lean on simulator-perfect state. The ruleset that matters removes it: no pose, no GPS, no map, and any external connectivity is a disqualification.

What it costs

Naive approaches crash or get disqualified. One early run was DQ'd for arming too early; another sat stuck 38 degrees nose-down for 19 seconds. Every failure got logged, diagnosed from telemetry, and encoded back into the stack.

The solution

Built the full autonomy stack: classical computer-vision gate detection, state estimation, spline trajectory planning, a 90 Hz geometric flight controller over MAVLink, and a reinforcement-learning agent for the simulation phase.

What I learned

Negative results are deliverables. Proving that forward velocity can't be recovered from monocular vision below 1 m/s at race range saved weeks of tuning a dead end and reshaped the design. And competition rules are requirements engineering: one early-start disqualification became DQ-safe arming logic baked into the code.