This comment[1] made it clear to me what exactly was going on. Also, on the author's GitHub, there's a link to a similar project,[2] which runs interactively in

This comment[1] made it clear to me what exactly was going on. Also, on the author's GitHub, there's a link to a similar project,[2] which runs interactively in a browser, making the concept a bit more accessible. Imagine you were dropped into a familiar setting, but were blindfolded. You know this place literally like the back of your hand, but you're only allowed to observe your location in a very limited sense -- say, every 10 seconds, you can reach out in the four cardinal directions to feel what, if anything, is around you -- but you can otherwise move around freely. Given enough observations, you can slowly establish your location precisely. What was bugging me is the use of a maze. Mazes are meant to be "solved" and this doesn't solve the maze -- in the sense of reaching a goal point -- but rather only allows the robot to localise itself within it. Given that the robot necessarily has an internalisation of the maze (a map), it might have been nice to see it simultaneously solve it, while getting its bearings, rather than just wiggling about, seemingly at random. [1] https://news.ycombinator.com/item?id=21878928 https://news.ycombinator.com/item?id=21878928 [2] http://www.claudeonthe.net/ai/particle_filter/particle_filter.html http://www.claudeonthe.net/ai/particle_filter/particle_filte...