It needs to be extremely accurate and precise, within a few millimeters, because:
Mmm. Over what working volume? How long are the links in this arm? One of the things with articulated robot arms is that, as they get bigger, their accuracy gets worse, simply because the positional error on each joint has a longer distance to work over -- doubling the the length of an arm link will double the error, assuming the error in the joint (motor error, gearing backlash, etc) remains the same. And that's before you add in things like increased gravity sag along the length of the link, increased mathematical rounding error in the DH model, thermal expansion... the list goes on and on.
The accuracy/precision question is a pretty big one. It's possible to get very good precision (repeatability) out of a an inaccurate robot. In my experience, an off-the-shelf major brand robot generally has an accuracy/precision ratio of ~10:1 -- that is, a 0.25mm repeatability works out to a 2.5mm accuracy. Assuming perfect Mastering, load data, modelling, etc. And that gets worse under various circumstances.
If you can afford to re-teach the robot at the critical points, and let it "run free" in open-air moves, you could null out a lot of these effects. OTOH, if you're running purely off a digital twin model and need to achieve high accuracy, even if you achieve Mastering perfection, you may have no choice but to re-calibrate the robot for each use case.
