1. Read what the standard pins down. ISO 9283:1998 Tables 1 and 2 specify drift of pose accuracy and drift of pose repeatability at 100 percent rated load and 100 percent rated velocity, with no optional second operating condition. Every other characteristic in those tables has one. Cell 1 lays the three rows out and counts the characteristics tied to a single corner.
2. Note what the extract does and does not carry. Secondary sources give an eight hour test duration for the drift test under clause 7.6. The preview extract the review retrieved did not contain it, so it stays a secondary-source figure and does not enter your arithmetic.
3. Price the envelope against itself. The same cell divides 50 mm of safe Cartesian worst case at stop by 0.1 mm of point repeatability. The 500 to 1 ratio is two accuracy claims about one arm inside one datasheet, which is what happens when a figure carries no operating condition with it.
4. Now the task side. Cell 2 loads Table 3.4 of one thesis: six columns of nuisance variation, twenty trials per cell, two policies trained on randomised simulation and on real data. Compute both means. You should land on 17.3 out of 20 and 13.2 out of 20.
5. Read the tablecloth column. 20 out of 20 against 1 out of 20, so 100 percent against 5 percent. That single column is the cleanest measurement of drift on the task the five-region sweep located.
6. Ask what the protocol can resolve. Cell 3 divides one trial by twenty. Every cell in that table moves in steps of 5 percentage points, so the object-colour difference of 5 points is one trial and is not a result. Four of the six cells clear the resolution.
7. Import the variance term, and label it. On another benchmark, with task, props and model held fixed, measured real success moved across the whole range from 0 to 100 percent depending on which class of human reset the scene. Read that spread as a caution about rankings rather than as the error bar of the table you just computed, and the cell labels it that way when it prints it.