/ Lab
A rocket that learns to land by itself
100 rockets, one neural network each, no instructions. Every generation the best ones pass their "brain" on to the next, with small mutations. After a few dozen generations they land, even with wind, rain and a floating platform.
Starting engines…
Simulation
Watching
Your simulation
rockets learning right now, from scratch
1
Gen · try 1/3
- Landed
- 0/100
- Weather
- Clear
- Wind
- 0 km/h →
- Difficulty
Rockets landed per generation
How it works
- 01
Real physics
Gimballed engine, side thrusters, fuel that burns off (the rocket gets lighter), engine lag, air drag. Contacts and impacts computed with Rapier.
- 02
One neural network per rocket
It sees position, speed, tilt, fuel and whether its feet are down. It decides thrust, engine angle and side thrusters, 60 times a second.
- 03
Evolution
Each generation flies 3 times in different conditions. The ones that land best (slow, upright, centred, with fuel left) reproduce. Difficulty rises when the best ones always land.