Implementing this successfully can be a nightmare.

Those are some 'gotchas' indeed! Particularly that middle one - it will take some integration between the smoothing and direction code to know to reverse but not turn until the robot velocity changes signs.

Lately, I've really been thinking about minimum-risk type movement. If you generate a bunch of movement options and then simulate them, it is easy to see whether they hit the wall. The trick is generating points which are likely to produce good movement options without having to generate millions of points.