Rover lab: drive a leg of the base
Pick two sites on the LOLA terrain, an autonomy level and the lighting, and drive the leg in your browser with the same model the logistics simulation's driving tables come from: slope limits, wheel slip, up-sun glare and shadow, Earth–Moon latency, energy and interventions. Drive it yourself to feel what a four-second round trip does.
Arrow keys: ← → steer, ↑ full throttle, ↓ stop. Your commands reach the rover one round trip later, and it halts to confirm every so often, exactly as in the model. Touch:
Loading…
Grid azimuths: at the pole “north” is the map's +y direction. Shadows are cast by terrain inside the tile only, at a 1.5° sun.
What the driver sees and does
A driving policy gets a 32 × 32 grid around the rover, 5 m per cell, rotated with the rover (forward is up; the outlined cell is the rover), plus five numbers. It answers with two: a steering rate and a throttle. These are the exact inputs the learned driver gets; when it drives, the wheel and the throttle below show the network's outputs.
The learned driver
A convolutional network that imitates the scripted expert from the grid and the five numbers above. It is fetched only when you choose it in the Driver menu (the network and the in-browser runtime are about 15 MB). The logistics experiments in the paper use the expert's driving tables, not this driver.
Against the driving table
The logistics simulation does not drive rovers step by step; it samples leg times, energy and interventions from tables built with this model. Here is the same leg three ways.
Constants
Every number carries its flag; hover a flag for its source.