a✦ALKAMY PLAY LAB FLIGHT EXPERIMENT 02

A LITTLE LIFT. A LITTLE LEARNING.

Skybound✦

Your favourite little robots, learning to fly.
Take the controls, or let a tiny neural pilot take the next flap.

170
learned parametersOne decision: flap or coast.
READY FOR TAKEOFF

MEET YOUR NEURAL PILOT

Ready, little robot?

Watch a locally trained model navigate the gaps.
Every flap comes from its learned weights.

Or choose “Fly yourself” to take the controls.
AI chooses every flap. No rescue guard.

Loading local flight weights…

SMALL BRAIN. REAL TRAINING.

A game you can play.
A model you can inspect.

The pilot learns from a simple flight teacher. Training, validation and testing use separate course seeds. Then the frozen model flies 100 fresh courses.

01 / LEARN

Flap. Coast. Repeat.

27,907 labelled states. A 170-parameter network trained locally for 800 epochs with NumPy.

99.5%held-out action accuracy4,613 states from 5 test courses

Imitation learning from measured game state. No pixels, LLM, reinforcement learning or hidden safety controller.

02 / FLY

Fresh courses. Same rules.

PilotMean gatesCrashes
Loading saved trials…

100 trials per controller. Each stops at the first crash, 50 gates, or 5,000 ticks. Scores are capped, not evidence of endless flight. The untrained pilot also clears this easy benchmark, so flight scores show no training gain.