This is an animation of the best aggregate map of 200 particles doing grid SLAM. It wiggles when another particle becomes the best aggregate particle. The map represents 10cm per pixel with a robot motion of 30cm per second. It is just dumbly wandering around, using a Vector Field Histogram to avoid walls - no purposeful entropy reduction.
It is using 24 simulated sonars modeled as fairly short range (200cm), narrow beam, medium motion noise and a beam probability model whose results are trusted to the power of 0.1 (i.e. not much, so as to encourage a smooth distribution without heavy particle duplication upon resampling - typically only 5%).
When the particles are moving into unknown territory, they diverge. When they see parts of the map a second time, they converge. As all of the particles' poses are overlaid on the current best particle's map, it can look as if they go through walls - if you were to see their own map, you would see that they think they are in the middle of the corridoor, and that the corridoor is in a different place.
You can see that the particle diversity is only just enough to close the loops.