With the versions currently available for downloading, the bots "sample" their field of view exactly the way humans do, and navigate in consequence. Actually I can boast to say they're the first and only bot to REALLY SEE their environment
however that doesn't make them the smartest mechs around, because it's a very CPU-intensive process and the bot is far from being terminated ; it was just to show people that it is possible to make a bot that doesn't use waypoints at all, a bot that is virtually able to play on every map you put it in - and that is already a cool thing
In the next release however (although it is already implemented in template #2 but not activated), the bots will make use of a kickass feature: the navigation mesh, that will allow them to LEARN the map. Cause currently, all they do is wander around, but they have no memory ; i.e. when they turn back it's just as if they thought they never walked that way before, and that's why you can sometimes see them doing stupid things. But with the navmesh working, and the independent time-sliced A* pathmachines, the bots will REALLY learn from your moves, and will be able to follow a player everywhere he goes. I have good hopes that this will be one of the l33test navigations around
A technical paper on how to do this is available at:
http://racc.bots-united.com/tutorial-navmesh.html
I'm currently in the last phase of this development, that is, the bots already learn the map pretty well, and they are already able to figure out paths to go everywhere, and I'm writing the function that makes the bots walk and follow the paths they've found. On this domain I'm a bit late compared to other bots such as RealBot and JoeBot XP who've done a great work already