Where does a trained drone controller break?

November 12, 2026

Method

Prove sweeps in simulation. Sixteen policies, a cascaded PID, a geometric controller and two references flew gusts from 16 directions at 11 speeds and two timings, 24 flights a cell on shared seeds, with an interval on every cell's crash rate and v50 fitted per direction: 174,080 flights with a re-check on fresh seeds.

Result

In simulation, sixteen RL policies trained in calm air crash in half their flights at gusts of 8.3 to 11.5 m/s, depending on direction: above a cascaded PID's mark in 8 of 16 directions and below a geometric controller's in all 16.

Could not detect

There is no test between a policy and a controller, because the analysis plan defines none, so the post says where each sits and not that one beats another. It is one figure-eight, gusts at two points on it, one air model and one simulated Crazyflie, and above about 6 m/s of airspeed the air model is an extrapolation.

Example

Where does my controller break?

Artifacts

code · results

When a gust hits a small quadrotor, the practical question about its controller is where it stops holding the path: from which direction, and at what wind speed. Wind-control papers often test a few wind speeds or scenarios and report tracking error or success rate in each. That tells you which controller did better in those conditions, but it doesn’t show where in direction and speed each one gives out.

So we measured that edge directly, in simulation. We trained sixteen RL policies to fly a simulated Crazyflie along a figure-eight in calm air, then flew them through gusts from 16 directions at speeds from 2 to 16 m/s, at two points on the loop. A cascaded PID built on the Crazyflie firmware’s structure and gains and a tuned geometric controller flew the same gusts on identical air, along with two references: a controller handed the true wind, and a privileged oracle, as one controller’s bound on where flying stops. With a re-check on fresh seeds, that’s 174,080 flights. A flight counts as a crash if the drone gets more than 0.5 m from the point it’s tracking (a 0.5 m tube that travels with the target), tumbles past the firmware’s tilt limit, or hits the floor, and for every direction we found v50, the gust speed at which half of the flights crash.

Across directions, the policies’ average v50 runs from 8.3 to 11.5 m/s, lowest for headwinds coming from below. It sits above the PID’s in 8 of the 16 directions, below it in 7 and level to two decimals in 1, and below the geometric controller’s in all 16. All 53,825 crashes on the main grid, across the twenty flyers, were the drone leaving the tube; none tumbled or hit the floor first. Before anything was drawn, every published number was reproduced from the raw flights by an independent re-implementation: a different AI model, given only the written definitions, wrote its own code and imported none of ours.

The rest of this post is how we got there and what it does and doesn’t mean.

Where this sits

Wind-robust drone control is an active area, and the better recent work measures carefully. In simulation, Tasim and Sun trained a quadrotor to estimate the wind from its own onboard measurements and fly with that estimate, and reported a 48% cut in tracking error against a wind-blind baseline. GustPilot paired an RL planner with a geometric INDI controller and tested it in 80 real-flight experiments on a 50 g quadrotor, against a DRL-PID baseline. RAPTOR trained one small policy to fly ten different real quadrotors, with wind disturbance among the conditions it was tested in.

Each of those compares controllers in the conditions its authors chose. A map of where each controller breaks, across direction and strength on identical air, is a different measurement, and it’s the one this post makes. In RL more broadly, Patterson and colleagues argue that the scale of today’s experiments often conflicts with getting proper statistical evidence, and Jordan and colleagues argue that rigorous benchmarking is often too expensive to do. The design here follows that advice where it can: sixteen training seeds with every failed seed reported, rules written before any gust was flown, the same air for every controller, and an interval on every cell.

The map

Averaged over the sixteen policies, v50 runs from 8.30 m/s for a headwind coming from 35° below level to 11.46 m/s for a level tailwind. The PID runs from 8.40 to 11.00 m/s over the same directions, and the geometric controller from 9.12 to 12.98 m/s. The controllers don’t share the policies’ ordering: both have their lowest v50 for a gust from the right side and off level.

Metres per second don’t carry from one air model to another, so every speed in the figures also carries a force ratio: how hard the gust pushes a level drone, as a share of the most sideways force full thrust can hold against. At 10 m/s that’s 0.41. At 12 m/s it’s 0.57. Nothing crashed at 5 m/s or below.

The sixteen policies don’t move as one. In 3 directions all sixteen sit above the PID and in 5 all sixteen sit below it, and in 15 of the 16 directions the best and worst policy are a metre per second or more apart.

Because the air is shared, every policy flight has a partner flown by each controller in exactly the same conditions. Against the PID, the policy alone crashed in 3,203 pairs and the PID alone in 3,136, out of 135,168. Against the geometric controller it was 9,857 to 46.

Every crash was the same kind

Leaving the 0.5 m tube was the first rule broken in every one of the 53,825 crashes. That makes the tube matter. A drone that leaves it isn’t broken, and the flight keeps going, to the end of the episode or the floor, so we stored when each rule was crossed and re-read the same flights under other limits without flying anything again.