NavFn plans paths which brush obstacles when in tight spaces.

[hershwg]

For instance, going through a 1 meter wide hallway with a 50 cm wide robot, the path should ideally go right down the center. As it is, the planned path will have the robot brush the near corner of the hall when entering and exiting, and potentially just slide right along the edge the whole way down the hall.

The incoming costmap has a sufficient inflation radius and cost function which falls off appropriately from obstacles, but inside NavFn::setCostmap() those cost values are rescaled per the needs of NavFn.

The problem is that the rescaling does not spread the input values nicely over the output range, it concentrates them at the max value. Therefore you get a wide plateau of high cost values around obstacles. In open environments with few obstacles, this is fine because it stays away from the plateau. In crowded environments, all the free map cells have equal costs on this plateau, so the planner just does strict shortest-path in this area.