A* is less useful when you're not omniscient, that is, testing if a cell is blocked has a sensing cost. I ran into this in a game application. To find out if a cell is obstructed, I have to do a ray cast at a few points in the cell, which uses resources. A* requires sensing a large number of cells to collect non-useful data, and if you have a big, mostly open space with some obstacles, like the real world, it does far too much sensing.
So I ended up with a variant on Pledge's approach to wall-following. Head toward the goal until an obstacle is detected. Then, start wall-following, but simultaneously in both left and right directions. When one of the wall-follower tests can head towards the goal, do that, and kill off the other wall-follower. So you alternate between heading towards the goal in open space, cheaply, and wall following.
Searching both left and right simultaneously avoids taking the long way round some obstacles.
So I ended up with a variant on Pledge's approach to wall-following. Head toward the goal until an obstacle is detected. Then, start wall-following, but simultaneously in both left and right directions. When one of the wall-follower tests can head towards the goal, do that, and kill off the other wall-follower. So you alternate between heading towards the goal in open space, cheaply, and wall following.
Searching both left and right simultaneously avoids taking the long way round some obstacles.