If you're using AStar, short of intricately defining the motion of each operator, my best suggestion is to change the acceleration and deceleration of the operators to 0. One big problem in this model is that the pickup prediction heuristic (the thing that's figuring out where along the conveyor to go to pick up the item, based on its position and speed) isn't working very well because acceleration/deceleration are significant portions of an operator's motion, which the prediction heuristic does not take into account. So the operators are predicting where to pick the item up, then going there, then realizing they're behind, so they then start chasing the item down the conveyor, and in doing that, they start walking over the center of the conveyor. By removing the acceleration/deceleration component, the operators make a much better prediction of where to pick the item up, so they don't start chasing it down the conveyor.
This solution won't make it perfect, especially when the operators still approach the conveyor at close to parallel. But it should make it better.
