Warehouse Robots Are Getting Better at the Task That Used to Stump Them: Picking Irregular Items

For years, the hardest problem in warehouse automation was not moving boxes, it was picking them: specifically, the large share of real inventory that is not a rigid, uniform box — clothing in poly bags, loose produce, irregularly shaped hardware, items in torn or crumpled packaging. Rigid grippers optimized for boxes handled these poorly, and human pickers remained the fallback for a large share of SKUs. The gap has narrowed significantly as two technologies matured together: tactile sensors sensitive enough to detect slip and adjust grip force in real time, and vision-language models that can plan a grasp point by reasoning about an item's likely shape and weight distribution from a photo, rather than needing a pre-trained model for every SKU shape. The practical effect shows up in pick-success-rate numbers: facilities that struggled to automate more than the "easy" 60-70% of SKUs are now reporting successful automated picking across a much larger share of catalog, including categories that were considered automation-resistant as recently as two years ago. Full lights-out fulfillment remains a longer-term goal rather than a current reality — damaged packaging, genuinely novel shapes, and safety-critical items still route to human pickers — but the boundary of what counts as "automatable" has moved substantially, and warehouse operators are re-running their automation ROI calculations accordingly.