Ask a fulfillment-automation engineer what breaks the business case and they won't say the AI. They'll say: the robot can't reliably pick up the weird object. Grabbing arbitrary items — a rigid box, a floppy bag, a dented can — at speed, without dropping or crushing them, is still unsolved hardware. The picking patents prove it.
Soft Robotics' grant US10518423B2, "End of arm tools for soft robotic systems," attacks it with compliant, soft grippers that conform to an object's shape — trading the precision of rigid fingers for the forgiveness needed to handle variety. X Development's US10518410B2, "Object pickup strategies for a robotic device," works the other end: given a pile of mixed objects, decide where and how to grasp. And Amazon's US10471599B1 even claims magnetic coupling for item manipulation — an admission that conventional grippers fall short on some loads.
Berkshire Grey's US10438034B2 zooms out to the whole station — "space efficient distribution stations and automated output processing" — because in fulfillment the grasp is one step in a sortation line, and the line's throughput is capped by the slowest, least reliable pick.
ROI per square foot, not per keynote. A warehouse can install all the planning software it wants, but if the end-effector misses one pick in twenty, a human has to babysit the cell, and the labor savings evaporate. The reliability of the physical grasp is the number the business case turns on — which is why so much picking IP is mechanical (gripper geometry, suction, compliance) rather than algorithmic.
Backlog is the only honest demo, and the picking patents are the unsexy reason backlogs convert slowly. Each new product category a fulfillment operator wants to automate is a new grasp problem; the gripper that handles books may fail on apparel. The hardware doesn't generalize the way software does, and the patents — spread across soft grippers, suction, magnetic coupling, and grasp strategy — show an industry still brute-forcing object variety one end-effector at a time.
What the filings can't claim is a pick-success rate across an open catalog — the single metric that would tell you whether general picking is solved. It isn't, and the breadth of competing approaches is the proof. For anyone modeling warehouse automation ROI, the gripper is the line item to stress-test, not the AI.