How it works
Seeing in 3D
The rover has no map to start with, only the edge of its operating area, and it assumes unknown ground is probably drivable. That optimism is what lets it plan at all, and also why it has to replan so often.
Everything it learns comes from three stereo camera pairs. The pair on the mast pans to look along the route and sweeps from side to side while the rover waits. Wide-angle pairs low on the front and back cover the ground around the wheels, which the mast can't see past the body. Each frame becomes a cloud of 3D points. The simulation makes the points by casting rays against the real surface and adding the error a stereo pair really has: it grows with the square of distance, so far points are much shakier than near ones. Turn on Sensors to see each camera's view traced on the ground and the points streaming in.
The points fall into a one-metre elevation map. For each cell the rover estimates the ground height from its lowest points, the tilt across its wheelbase from the cells around it, how bumpy the ground is, and whether something stands up out of it, like a rock. It never sees the true terrain, only these estimates, so the back of a boulder stays unknown until it drives past, and far ground can be misjudged until it gets closer.
Knowing its own limits
One vehicle description drives everything: a 3 by 2.2 metre body, wheels that can climb about 34 centimetres, a 20° limit on the slopes it will choose, and 27° before it risks tipping. Ground past a limit is no-go, hatched on the heat map. That is where a wheel must never go. Around every no-go cell there is a keep-out ring about half a body wide, where the centre of the rover can't go because its side would be over the hazard. Beyond that is a ring where a turn could swing a corner over it, which is costly but allowed. Steps are flagged a little below the true limit, as the estimates are noisy and a missed rock costs far more than a detour.
An IMU reports the body's roll and pitch, shown under the status. The rover slows on slopes it can feel. If the tilt nears its limit it stops and backs off, whatever the map said.
Planning
Routes come from A* on an 8-connected one-metre grid, then get straightened wherever a direct line is clear and no more costly. The search runs in a Web Worker so it never stalls the animation. When the cameras find a hazard on the route, the old path flashes orange and a new one is planned; smaller changes are folded in quietly while it keeps driving. The cost map updates only the cells new points touch.
Staying out of no-go zones
Every route is checked against the latest map before the rover drives it, so a plan worked out before a new hazard appeared is thrown away rather than followed. While driving it refuses any move that takes its centre into a keep-out ring or puts a wheel on no-go ground it knows about. When that happens it follows the grid route closely for a few metres instead of cutting corners. If it finds itself inside a keep-out ring, because a hazard was only just seen, it picks its way out along the direction that gets clear fastest.
When things go wrong
The body also checks its footprint against the real ground before every move, which catches what the cameras got wrong. If it touches a rock, the tilt gets too steep, or it stops making progress, it marks the spot as no-go, reverses along its own tracks (the one path it knows is clear), and replans. After four failed recoveries in a row, or if no route exists through what it knows, it stops and says it is stuck. Some places are simply stuck.
What's simplified
- Stereo depth is simulated by ray casting with range-dependent noise, not by matching real images, and the rover always knows exactly where it is. A real rover would also estimate its own position from wheel odometry, the IMU and visual odometry, and that estimate would drift.
- Every replan searches from scratch. That is fine at this size; an incremental planner like D* Lite would reuse earlier work on larger maps.
- The suspension is animated to follow the ground, not simulated as a mechanism.
The terrain is a generated lunar surface: rolling regolith and a power-law crater field, lit by a low sun as near the poles. Lighting and shadows for the ground are computed once in a Web Worker along with the terrain, so the page only draws; the rover is the one thing with a live shadow. It is a toy built from textbook methods and a generic design, and none of it downloads until you reach for the Run button or scroll to the model.