If this is the same problem I was having, it was reduced by setting the object type to PH_POLY -- but not totally gone.

It also seemed to me that there are particular locations where the object was more likely to pass through the wall than others. I think I read something by ventilator about this, saying, "he knew where to look" for the problem.

I have not had good luck with setting correction values higher because this seems to contribute to the problem of the physics object reacting violently to vertex and edge collisions.

Therefore, I have had best results with PH_POLY and low corrections, but I arrived at this by trial and error and cannot really explain whether this is an appropriate way to do physics.