Teaching a robot where to pick from a pile of grapes

We've been developing a vision capability for robotic picking of loose food product — the sort of task behind pre-packed salads, sandwiches, and fruit punnets.

The video shows the vision system in isolation, not the robot or gripper. Its job is to work out, from a jumbled punnet, which item to pick next and exactly where to place the pick point.

For grapes the logic is straightforward: build a height map of the punnet, find the highest point, and mark it as the pick point. Take the top grape away and the pile settles predictably for the next pick — no need to re-plan for what's underneath until it becomes the new highest point.

Simple in concept. It's also the baseline for two much harder versions of the same problem — strawberries, where the pick point has to avoid the calyx, and chicken breast chunks, where true piece size can't be known and the system has to work with what's visible.