It indeed was not clear about the specific problems. It basically implied that it shouldn't pick all strawberries, just a subset, probably the most ripe. But it didn't describe well what the bot keeps getting wrong.

I suspect people use tactile feedback in addition to sight, but that is difficult for bots to get right because it requires lots of contact sensors, lots of processing, and lots of coordination with vision to tune well (at our current stage of AI). But it does sound like it's getting incrementally better such that the bots' time will eventually come. I wonder if a spectragram could reduce the need for tactile sampling? That's something a bot has up on humans.