SOLO — Teaching a Neural Network to Fly, From Scratch
I wanted to watch an AI become independent — literally — so I taught a neural network to fly, then cut the cord.

What it is
SOLO is a real neural network, hand-written in one file (forward pass, backprop, and a full neuroevolution loop — no libraries), that learns to fly a gated obstacle course. It can learn by cloning your piloting (with DAgger coaching to correct its drift), or by teaching itself from scratch with no teacher at all. Then it flies a course it has never seen, and you can watch the network fire live as it does.
How it’s built
- A from-scratch MLP with forward + backprop + SGD with momentum; output[0] steers via tanh, output[1] throttles via sigmoid.
- Two learning paths: behavioral cloning (+ DAgger) from demonstrations, and neuroevolution (population, fitness = flight progress, elitism + tournament + crossover + gaussian mutation) with no teacher.
- Verified on the honest metric — completion of an unseen course (~91% cloning, ~95% evolution), not a misleading input-match proxy.


On the theme
Independence is a measurable moment here: when the network stops needing you and completes a course it’s never seen, on its own.
Try it →Code →All the builds →
Shout-out
Part of the competition is cross-referencing other builders. So: shout-out to Eric Rhea — host of Summer Into AI; worth a follow if you’re tracking the competition.
Built for Summer Into AI 2026 (Competition #2), hosted by Eric Rhea. More in the build log.