The floor at the controls
of a drone.
A recorded run of the flight console, replayed panel for panel. The drone flies the route it learned by trial and failure on the course from the open-source jev-drone project. The learned plan sets the speed; the stop-distance limit and the reflexes are applied after it, so they can only slow it down. There is no neural network anywhere in the loop.
What you are watching
- The view. third-person, from the simulator (MuJoCo, Skydio X2 model).
- Onboard camera · depth. the only sensor the rules read.
- The route it has learned. one tile per 6 m block: the speed cap it earned and whether it flies high there.
- Learning curve and rounds. the 87 rounds that built this route: every change it tried, kept only if every lap came back cleaner and faster.
- Situation report. exactly what the rules read on this frame: free space by sector, the path ahead, the nearest obstacle.
- What happened. the lap log. In this recording every lap finishes in 27.6 s with 0 contacts.
Watch it learn from nothing
The replay above flies a route that is already learned. The recording below starts from a wiped memory: seven rounds take the cost from 69.39 to 51.52, two changes kept and five reverted. Contacts happen while it learns; each costs 15 s and a round that adds them is thrown out.
Results, with the caveats
On held-out starting positions, run once after the rules were frozen, the floor flew the whole course 10 of 10 times with 0 contacts. The unmodified baseline finished 1 of 10.
Caveats that always go with that number: the course is identical in every run and only the start pose varies, so this is weak evidence of generalisation; it is simulation, not flight; the upstream project is pinned and unmodified, and we did not re-run anyone else's system.
How it connects
This is the same idea as a fail-safe model: the learner changes what the drone tries, never what it is allowed to do. The decision path is a symbolic flow: engine state becomes facts, the rules choose, the actuator acts. The flight floor itself is described on Flight and The engine. Peel is, to our knowledge, the first fail-safe model.