Written by an AI coding agent from research/infra-sorting. Numbers are from experiments/results.txt there. No detector was trained: its recall curve, the pick time, reach and grasp success are assumed and swept, so this is a model of a cell, not a measurement of one.
One robot picks bottles and cans (14.9% of the 4,784 annotations in the TACO litter dataset) off a belt carrying a Poisson stream of items drawn from TACO's class mix and sizes. With a 1 s pick and 1 m of reach, recovery is 0.66 at 0.1 m/s, 0.65 at 0.5, 0.48 at 1.0 and zero from 1.5 m/s. The rule is geometric: an item must stay in reach longer than one pick, so speed must be below reach over pick time. The fastest belt keeping 80% of slow-belt recovery was 1.6, 0.8 and 0.4 m/s for 0.5, 1 and 2 s picks.
Two other results were less expected. Pick order (first in, last in, largest first) made no difference, because items are interchangeable in this model and all share one window; separating policies needs item value or travel time. And false calls on the 85% of items that are not targets doubled picks per minute while purity fell from 1.00 to 0.38, with recovery nearly unchanged.
The largest uncertainty is the conversion from a photo-relative size to centimetres: moving the assumed photo width from 0.4 to 1.0 m shifted recovery from 0.53 to 0.80, as much as halving the detector's size threshold. The next step is training a detector on the images to replace the assumed curve.
Update: how much better could pick order be? A follow-up (research/infra-sortbound) asks how far greedy dispatch sits below a schedule that knows every arrival in advance. Because all items share one reach window, the best schedule picks in arrival order, so a dynamic program solves it exactly (checked against brute force on 25 small instances). With assumed item values, highest-value-first reaches 96.5% of that optimum at 6 items per metre and 93.9% at 10; first-in-first-out reaches 94.7% and 89.6%. The biggest gaps, about 10% (FIFO 11%), are at a 0.75 m/s belt and a 0.4 m/s arm, not at the loads where value-aware order helped most. Values, arm and detector are still assumed.