The trick is to detect obstacles that are likely to have minimal affect on the global result of pathfinding search, e.g. convex obstacles that are in the middle of open spaces, or, more specifically in the case of PathEngine, obstacles that don't combine with other nearby obstacles to form larger blockages.
After the initial graph search, the path is modified to avoid these obstacles locally (by pushing the path around the obstacles, essentially), before it's returned from the pathfinding query.
So the way it's set up in PathEngine this optimisation is largely hidden from application code (code that calls into the pathfinding API).
Leaving this to be handled later on, by agent local obstacle avoidance, could also work..
After the initial graph search, the path is modified to avoid these obstacles locally (by pushing the path around the obstacles, essentially), before it's returned from the pathfinding query.
So the way it's set up in PathEngine this optimisation is largely hidden from application code (code that calls into the pathfinding API).
Leaving this to be handled later on, by agent local obstacle avoidance, could also work..