Mapping

A map is a memory the robot is willing to reuse. If you only need to avoid the next two meters, you may not need one. If you need to return to a charger through a building, you do. Start with purpose. A pretty point cloud that nothing plans on is a souvenir.

Occupancy grids are the workhorse for indoor mobile robots. Space is tiled. Each cell stores a probability of being occupied, often in log-odds so updates are additions. A range hit increments occupancy along the endpoint; the free ray decrements the cells it flew through. Unknown is a first-class state — do not treat white as free just because you never looked. That mistake is how robots invent hallways.

Topological maps remember places and the doors between them: “kitchen connected to hallway.” They are cheap to talk about and brittle if the “place” detector is vague. Metric maps remember coordinates and can support A* on a grid. Grown systems often hold both: a grid for local motion, a graph for “which room.” Feature maps (landmarks with poses) sit closer to SLAM.

Mapping assumes you know where you were when you measured. That assumption is the localization chapter’s problem. If you stitch scans with raw odometry, the map smears. If you pretend the map is finished and then localize, you are doing sequential SLAM by stages. Be honest about which assumption is in force today.

Lab: push a lidar down a corridor in a straight line you measured with a tape, insert cells, and print the grid. Then add a small yaw bias and watch the corridor bend. That bent corridor is why we bother with Part IV. Read the book chapter for the update rules and representations this note only sketches.