Industries / Transport and vehicles / Warehousing and distribution
Picking
Picking is taking the items for each order off the shelves, whether a person walks the aisles with a list, a robot brings the shelf to a person at a station, or a robot arm lifts each item.
7
Milestones achieved, from 1906 to 2024.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1906
Orders run on a timetable
Sears, Roebuck opened a mail-order plant on Chicago's West Side with more than 3 million square feet of floor space. Each incoming order was given a time to be shipped, and its goods had to reach an assigned bin in the assembly room by that time, carried there by "an intricate system of belts and chutes." Sears says the system became "a sort of 'seventh wonder' of the business world."
Credit
Sears, Roebuck mail-order executives (the Sears archive names no single designer)
1974
A code a machine reads
A pack of Wrigley's gum became the first product scanned with a barcode, at a Marsh supermarket in Troy, Ohio, on 26 June 1974, according to GS1, the standards body that runs the code. By 1983, GS1 says, the same codes had spread beyond checkouts to multi-packs, cases and cartons in wholesale, giving each item and case a number a scanner can read.
Credit
US grocery industry leaders and the Uniform Code Council (now GS1 US)
1997
Picking by voice
Vocollect, founded near Pittsburgh in 1987 by former Westinghouse engineers Roger Byford, Mike Gabrin and Larry Sweeney, built a wearable unit with Telxon, the Talkman, that speaks each instruction to the worker and hears the spoken reply. It replaced paper pick lists and left both hands free, and by 1997 six Wal-Mart distribution centers were using it, according to a 2002 profile.
LessonVocollect first sold voice entry to factory inspectors and nearly failed; it found its market after a founder visited a hardware distribution warehouse.
Credit
Vocollect; Telxon; Wal-Mart as an early customer
2012
Shelves that come to you
Amazon agreed on 19 March 2012 to buy Kiva Systems of North Reading, Massachusetts, for about $775 million in cash. Kiva's wheeled robots slide under a shelf, lift it and carry it to a worker at a fixed station, so the picker no longer walks the aisles; this is called goods-to-person picking. Amazon said on 30 June 2025 that it had deployed its one millionth robot, by its own count.
Credit
Kiva Systems (founder and chief executive Mick Mountz); Amazon
2015
A shelf-picking contest
At the first Amazon Picking Challenge, held at the ICRA robotics conference in Seattle, more than two dozen teams built robots to pick 12 named items from a warehouse shelf within 20 minutes. Team RBO of Technische Universität Berlin won with 148 points, picking 10 of the 12 items.
LessonThe organizers' analysis put the best robot at about 30 items an hour with a 16% failure rate, against roughly 400 an hour for a person, and teams ranked seeing and grasping the items as the hardest parts.
Credit
Amazon (organizer); Team RBO, TU Berlin
2019
A learning arm in production
In September 2019 a robot arm supplied by Knapp, an Austrian warehouse automation company, and run by software from Covariant, a startup in Berkeley, California, went into full production at a warehouse of Obeta, a German electrical wholesaler, picking single items from bulk totes (plastic bins) into order boxes. Knapp said in January 2020 that the arm could pick about 95% of Obeta's product range, up from 10 to 15% before, including shiny items and items in plastic bags; the rest, mostly fragile goods, stayed with people.
Credit
Covariant (Peter Chen, Pieter Abbeel); Knapp (Peter Puchwein); Obeta
2024
A humanoid robot's paid job
GXO Logistics, a contract warehouse operator, signed a multi-year agreement with Agility Robotics in June 2024 to use Agility's two-legged Digit robot at a GXO site serving Spanx, moving totes from mobile robots onto conveyors. GXO called it "the industry's first formal commercial deployment of humanoid robots." Digit had moved more than 100,000 totes at the site, in Flowery Branch, Georgia, by November 2025, a figure the trade magazine DC Velocity attributes to GXO.
LessonThe paid work is moving whole totes, not picking single items out of them.
Credit
Agility Robotics (chief executive Peggy Johnson); GXO Logistics
Watch for
Amazon's arm with touch
Amazon says Vulcan, a robot arm with force sensors that tell it how hard it is pushing or gripping, can pick and stow about 75% of the item types stored in its fulfillment centers at speeds comparable to its workers. It works the top rows of storage pods, about 8 feet up, at sites in Spokane, Washington, and Hamburg, Germany. Amazon gives no count of Vulcan robots or picks. Amazon claimed the same 75% for its Blue Jay multi-arm system in October 2025, then added a note on 25 February 2026 that it was no longer using Blue Jay in operations.
Achieved whenSomeone outside Amazon reports Vulcan in routine use at a named site, with figures such as the share of picks it handles.
Claimed by
Amazon, May 2025
600,000 hires not needed
Internal Amazon documents reported by The New York Times say automation could let Amazon avoid hiring more than 600,000 US workers it would otherwise need by 2033, while it roughly doubles the goods it sells, and avoid about 160,000 hires by 2027. Its robotics team reportedly aims to automate 75% of operations. Amazon said the documents do not reflect its full strategy.
Achieved whenAmazon's own filings or an outside count show the units it ships about doubled by 2033 while its US warehouse workforce stayed near its 2025 level.
Claimed by
Amazon internal documents, as reported by The New York Times, October 2025
Digit at two more sites
GXO plans to put Digit to work filling e-commerce orders at its sites in Statesville, North Carolina, and Lebanon, Tennessee, with deployments expected to begin in early 2027. No robot count was given.
Achieved whenGXO reports Digit in routine work at those sites, with tote counts or hours worked.
Claimed by
Agility Robotics and GXO, October 2026
Open
01
Hard-item picking tally
A low-cost robot arm that picks the items warehouse robots still struggle with (clothing in plastic bags, items in clear plastic, soft and floppy goods) out of a tote one at a time, with every attempt filmed and logged as a success, a double pick, a drop or damage.
The briefCredit · unclaimed02
A full-shift pick log
A test in which a picking robot runs for a full eight-hour shift, or 1,000 picks in a row, on a tote of mixed items, with a timestamp and outcome for every attempt, so that picks per hour and error rate are measured in the units vendors quote, under a written protocol any other team can rerun.
The briefCredit · unclaimed
Open 01 · the brief
Hard-item picking tally
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A low-cost robot arm that picks the items warehouse robots still struggle with (clothing in plastic bags, items in clear plastic, soft and floppy goods) out of a tote one at a time, with every attempt filmed and logged as a success, a double pick, a drop or damage.
Why it is still open
The largest public warehouse picking dataset, Amazon's ARMBench, with more than 235,000 pick-and-place records, covers seeing and identifying items, not whether picks succeed. Lab benchmarks use lab object sets, and the published results found for low-cost arms are for cubes and blocks. No public attempt-by-attempt tally for these hard item classes turned up.
Built with what exists
An SO-101 arm with Hugging Face's LeRobot software (see the Note "No download"); an open robot model such as Runway Praxis-1 (see the Note "Open robot model weights") or NVIDIA GR00T 2, which learns a task from a few demonstrations (see the Note "Five demonstrations per task"); a parallel gripper or small suction cup; a webcam.
First step
Buy ten polybagged clothing items, record 50 hand-guided demonstrations, run 20 attempts and publish the first tally.
Done looks like
30 items, ten in each of the three hard classes, with 20 filmed attempts per item (600 in all), published with video, a spreadsheet of outcomes and the success rate per class.
Open 02 · the brief
A full-shift pick log
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test in which a picking robot runs for a full eight-hour shift, or 1,000 picks in a row, on a tote of mixed items, with a timestamp and outcome for every attempt, so that picks per hour and error rate are measured in the units vendors quote, under a written protocol any other team can rerun.
Why it is still open
Public contests and benchmarks measure short runs; the 2015 Amazon Picking Challenge gave each robot 20 minutes. Vendor figures, such as Knapp's 95% of range, Nomagic's 300 packs an hour and Amazon's "speeds comparable" to workers, come without attempt logs. No public multi-hour log of picking attempts turned up.
Built with what exists
An SO-101 arm with LeRobot (see the Note "No download"); an open robot model (see the Note "Open robot model weights"); a camera; a script that timestamps each attempt; a reviewer model that flags doubtful outcome labels for a person to check (see the Note "Seven more flagged").
First step
Write the protocol (the item list, and what counts as a pick, a double pick and a drop), then run and publish one logged hour.
Done looks like
1,000 attempts in a row on 50 mixed items, with video, a public log file, picks per hour, and the error rate with its margin of error, plus the protocol document.