Rocket Lander

This rocket is flown by a small reinforcement-learning agent, trained from scratch by letting it crash a few hundred thousand times. Change the world and watch it cope. It was trained on a range of gravity, wind, and pad positions, so it handles those well; push a slider past where it ever trained and you can watch it struggle. You can also drag the rocket to fling it, or drag the ground to move the pad.

With autopilot off, fly it yourself: ↑ thrust, ← → rotate.

Original code and physics. The task and controls are inspired by the classic LunarLander environment from OpenAI Gym, now maintained by the Farama Foundation as Gymnasium.