Game Lab · Evolution

Rocket Lander

Land the rocket gently on the pad! Hold ↑ (or Space) for the engine and ← → to tilt. Then switch to Robo learns and watch 60 rocket brains figure it out.

Hold ↑ for engine, ← → to tilt
Each generation champion's score everyone's average challenge on

Checkpoints — the champion brain over time. Tap one to see how it copes with today's world.

How it works

Feel

Each rocket feels 6 numbers: where the pad is, how high it is, how fast it's moving, and how tilted it is.

Decide

A tiny brain turns those numbers into three choices: engine on? tilt left? tilt right?

Score

Each brain tries 4 landings. Close to the pad, slow and straight scores points. A soft landing scores lots.

Breed

The best brains are mixed together and given small random changes to make the next generation.

Adapt

When you add a challenge, old tricks stop working. The ones that cope best get picked, so the family adapts.

What just happened?

For grown-ups: what's really going on

This is neuroevolution: a genetic algorithm searches the weights of a small neural network, with no gradients.

  • Policy: a fully-connected net 6 → 8 (tanh) → 3. Inputs: pad offset, altitude, vx, vy, tilt and pad velocity. Outputs: engine (fires if > 0), tilt left / right (if > 0.3 and larger than the other).
  • Fitness: mean over 4 landings per generation (shared start conditions, so the comparison is fair) of proximity + low impact speed + uprightness, plus a large bonus for a soft upright landing on the pad and a small bonus for fuel left.
  • Selection: the top 12% are elite; children use uniform crossover of two elites and Gaussian mutation on 25% of weights. The top two are copied unchanged.
  • Non-stationary environment: challenges change the dynamics (wind, gravity, fuel budget) or the task (moving / random pad). Fitness of the old champion drops, and selection pressure re-optimises the population. This is a toy version of how evolutionary strategies and RL agents handle distribution shift.
  • Note: the rockets can't feel the wind directly; they only notice it through their own drift (vx), so adaptation has to happen through evolution, not perception.

Related reading: the classic LunarLander RL benchmark, and OpenAI's Evolution Strategies article.