Lander lab: the last 500 metres
A lander's terminal descent from the SPLICE braking gate to touchdown, simulated in your browser by the same equations as the Python environment. Fly it with the trained neural-network policy or the classical guidance law, watch the score it was trained on, change the reward rules, and compare both controllers over 500 dispersed landings.
Loading the policy…
Side view along the approach direction. Solid line: this run; dashed: the other controller on the same episode. The ground profile is the LOLA terrain through PAD A for context; the environment itself lands on a flat plane with a sampled site slope (0–5°). The lander is drawn far larger than scale.
Reward rules
A reinforcement-learning policy learns whatever the reward pays for. Move the weights to re-score the episode you just watched; the table shows how much each term contributed. Presets are the variants the training code knows.
Monte Carlo landings: learned policies against the classical law
Each controller flew the same 500 dispersed episodes (same initial conditions, hardware spread and sensor noise), computed by lunarsim.lander.evaluate in Python. These tables are what the logistics simulation samples landing outcomes from.
The policies
Constants for this class
Every number carries its flag; hover a flag for its source.