It is a reliable surprise that a robot can weld a car body to a fraction of a millimetre all day, and struggles to take a random item out of a crate. The first task is repetitive, identical each time and performed on something held in a known position. The second requires working out what the object is, where it can be held, how much force it will tolerate and what else will move when it is lifted - decisions people make without noticing they are making them.
The sense of touch is the part machines lack most. A person picking up a paper cup adjusts grip continuously, feeling slip and compensating before the cup deforms, and does this for an object they have never held. Reproducing that needs sensing, interpretation and control operating together in a fraction of a second, and while it has improved substantially, it is still far short of a hand. Grip too gently and the item drops; too firmly and it is crushed. The margin is narrow and specific to each object.
Variety compounds it. A warehouse might hold a hundred thousand different items - rigid, soft, heavy, fragile, in bags, in boxes, loose - and a system that handles the first ninety percent still needs a person for the rest. Since the exceptions are unpredictable, the person cannot be removed; they can only be given fewer, harder cases. This is why so much of warehouse automation moves shelves to people rather than replacing the picking itself, which was the more practical thing to automate.
Where the objects are uniform, the problem largely disappears, and that is where these machines are in real use: identical products, known shapes, consistent packaging. The capability is not general dexterity but a narrow competence that happens to match a specific, valuable task. Understanding this distinction is what separates a realistic assessment of a machine from a demonstration video.
It also explains where the automation frontier actually sits. The jobs that resisted machines are not the ones needing strength or precision - those went first. They are the ones needing hands and judgement together in unstructured places: plumbing, repairs, care, cooking in a small kitchen, anything done somewhere slightly different each time. Those remain among the hardest things to build, and the gap between them and a robot arm in a cage is not closing quickly.