Robot piece picking: what customers actually confirm about pick rates

Customers of robot piece-picking systems rarely publish a pick rate a buyer can use. The clearest example is Zalando and its supplier Nomagic. In October 2025, Nomagic said the Zalando pilot achieved "an average of 10 000 picks/day." In March 2026, Zalando said the pilot "achieved a staggering 100,000 picks per day." In May 2026, Zalando said robots automate 2 million picks a month. None of the three figures says how many robots, how many hours or which sites it covers.

That missing denominator, per robot or per fleet, per hour or per day, average or best day, is the pattern across this analysis. The operators with the best data, Ocado and Amazon, publish something different again: site-wide labor productivity or the share of item types a robot can handle, not a robot's pick rate. In the sources reviewed here, a per-hour figure for a single robot appears only on vendor product pages. For a buyer, a pick-rate claim is unconfirmed until someone states the unit, the time window, the hours run, the item mix and how much human help the robot gets.

Analysis based on company releases, product pages and annual and half-year reports, all retrieved November 9, 2026. AI assistance was used in research and drafting, and the piece was checked by a separate AI fact-check and editorial review. Servospace did not visit any site, interview the companies named, or see operating data. Calculations are Servospace arithmetic on published figures and are labeled as such.

In this analysis

• What piece picking is, and what gets counted
• Zalando and Nomagic: one pilot, three figures
• Can the figures all be true?
• What the vendor's own spec sheet says
• Announced, selected, live: the status ladder
• What Ocado and Amazon publish instead
• Counterevidence and limits
• What a buyer should ask for

What piece picking is, and what gets counted

Piece picking, also called each picking or item picking, is a robot arm taking one item at a time out of a bin or tote and placing it somewhere else: into an order tote, onto a conveyor or into the pocket of a sorter. It is the step after storage. In a cube-storage system such as AutoStore, a robot arm can stand at the port where bins arrive, which is the station covered in our AutoStore installed-base analysis. Earlier steps in the building, the trailer dock and case handling, move whole boxes rather than single items.

"A pick" sounds like a fixed unit. It is not. The same word covers at least four different jobs, and a rate for one says little about another:

Rates also depend on the item mix. A rigid box picks differently from a polybagged shirt or a shoebox with a loose lid, which is the problem Nomagic says its Shoebox Picker solves.

Zalando and Nomagic: one pilot, three figures

Zalando, which describes itself as the leading European technology platform for fashion and lifestyle, uses robots from Nomagic, a warehouse robotics company. The two companies have published these statements about their work together:

Both releases describe the same job: item-level picking, scanning and induction into automated pocket sorters. That is sorter induction, not picking customer orders into a box.

The gap between the first two figures is tenfold, and the qualifier drifts between hops. Nomagic's figure is an "average." Zalando's carries no "average" and adds "staggering." Neither states the number of robots in the pilot, the sites, the operating hours or the period measured. The Nomagic release also discloses that Zalando made a minority investment in Nomagic as part of its $44 million Series B, so the customer is also a shareholder of the supplier.

Zalando's own financial reporting adds one figure and no denominators. A text search of its 2025 annual report finds no mention of Nomagic or robots, only a general line about AI "automating warehouse operations," and says more than 7,000 employees work in its fulfilment centres. Its first-quarter 2026 results release of May 6, 2026 does name Nomagic: "Already today, 2 million picks are automated by robots every month." The release does not give the robot count, the sites, the hours, or whether the 2 million is Nomagic robots only. Zalando's second-quarter and third-quarter 2026 results releases do not repeat a pick count.

Can the figures all be true?

Possibly, and the arithmetic shows how. If Nomagic's 10,000 picks a day was a per-robot average, and Zalando's 100,000 covered the "first ten units," then 10 robots times 10,000 picks gives 100,000. That reading would make the two releases consistent. Neither release says so.

The vendor's own speed ratings set limits on the readings. Nomagic lists 450 to 600 picks per hour for its Shoebox Picker (see the next section). Servospace arithmetic on those figures:

• 10,000 picks a day from one robot needs about 17 hours at 600 picks per hour, or about 22 hours at 450. That is possible only in near 24-hour operation, which large fashion warehouses can run, but no source states the pilot's hours.
• 100,000 picks a day from one robot would need about 4,170 picks an hour over 24 hours, roughly seven times Nomagic's highest listed rate. So the 100,000 figure cannot describe one robot at listed speeds. It must be a multi-robot or multi-site total.
• 10,000 picks a day across the whole pilot fleet of nine or ten robots would mean about 1,000 to 1,100 picks per robot per day, around two hours of work at 500 an hour. That is plausible for an early pilot, but then the 100,000 figure would describe a different, larger scope.

Zalando's May 2026 figure adds a cross-check. Servospace arithmetic: 2 million picks in a 30-day month is about 67,000 a day. That is below 100,000 a day, so if both figures cover the same robots, the 100,000 cannot have been a sustained daily rate; it reads more like a best day or a different scope. Neither release says which.

What the public record supports, then, is narrow: Zalando attributes 100,000 picks a day to the pilot and 2 million automated picks a month to its robot roll-out, with no robot count, hours or site list on either figure. It does not support a per-robot rate, a sustained rate, or a comparison with human pickers.

What the vendor's own spec sheet says

Nomagic's product pages give per-hour figures, with some conditions:

These are the most specific numbers in this analysis, and they come from the seller. They are ceilings ("up to") with no stated period, item list or test method. The Shoebox Picker page shows the label drift within a single page: the headline says 600, and the shoebox-specific figure is 450, 25% lower for the item the product is named after.

The pages also mark where people still sit in the loop. Nomagic describes an AI monitoring system and "expert remote operators" watching robots 24/7, a 15-minute response to requests and a team member on site within 24 hours if something is wrong. "Eight hours without human intervention" suggests that intervention does happen on longer runs. Neither page gives an intervention rate, a share of picks handed to people or a count of remote staff per robot. Remote help is common in this industry: Tutor Intelligence, for example, has human "tutors" guide its robots remotely. A pick rate that does not say how many picks needed a human is incomplete.

Announced, selected, live: the status ladder

Most robot piece-picking news describes a stage before steady operation. The documents reviewed here fall on a clear ladder:

In every customer quote in the Nomagic releases reviewed, the customer speaks qualitatively ("an important step," "the next logical step"). The numbers come from the vendor, except for Zalando's own 100,000-a-day and 2-million-a-month figures.

The Otto Group case shows why the stage matters. In 2023 the Otto Group announced a partnership with Covariant that would "begin by deploying over a hundred robots," starting at its Haldensleben and Altenkunstadt sites in Germany. Modern Materials Handling reported, citing Covariant, that the first eight robots would be in Haldensleben and two in Altenkunstadt by October 2023, and 120 across Otto's network over four years. In January 2024, Otto's own story from Haldensleben described a single pilot test station installed in November 2023: "It is still the only one there." More stations would follow "if it passes all the tests," with testing on real customer orders still to come and regular operation at other sites targeted "as early as 2025." The Otto Group's annual report for 2024/25, published in May 2025, says that "after extensive tests with the AI station in Haldensleben, the AI robots will be gradually rolled out in all of the Otto Group's logistics operations." That is still a plan, not a count of live stations. Searches for this analysis found no later Otto Group statement of how many Covariant stations are running. A secondary article claimed more than 100 robots and a rate of 1,600 items an hour, citing a Covariant case study; no matching Otto or Covariant primary source was found, so those figures are not used here. In August 2024, Amazon licensed Covariant's robot AI models on a non-exclusive basis and hired its founders and about a quarter of its staff, saying Covariant would "continue to serve its dozens of customers."

None of this shows the Otto program failed. It shows that a 100-robot plan and a one-station pilot were both true at different times, and that the customer later described tests completed and a rollout still to come, without publishing a station count.

What Ocado and Amazon publish instead

The two operators with the deepest robot picking programs in this sample publish their results in a different unit. Both are also the technology suppliers, so their figures are company statements about their own systems, not independent customer confirmation.

Ocado reports labor productivity, not robot speed. Ocado's On-Grid Robotic Pick (OGRP) is a robot arm mounted on its storage grid that picks grocery items and packs them into bags. Ocado's half-year report for the 26 weeks to May 31, 2026 says OGRP is rolled out across 14 customer fulfilment centres (CFCs) for eight partners, with more than 350 million picks to date. Its full-year 2025 results say its most advanced CFC was "now picking c.50% volumes robotically," and its technology page identifies that site as Luton, UK, with a 45,000-SKU range.

Ocado's headline operating metric is UPH, which it defines as "average units picked per labour hour," measured wall to wall across a site. Luton averaged 289 UPH in the year to November 30, 2025, with a peak of 318, and 315 in the first half of 2026, up from 277 a year earlier. Ocado attributes the gains partly to OGRP and to Auto Frame Load, which automates loading packed totes into delivery frames.

That is the most useful kind of number for a buyer: it measures what the building produces per hour of paid human labor, after robots, exceptions and supervision. It has its own limits. It blends two automation systems with other changes, so it does not isolate the robot arms. And Ocado owns 50% of Ocado Retail, which operates Luton, so it is not an arm's-length customer report.

Amazon reports coverage, not rate. Amazon's Vulcan page says the robot can pick and stow "approximately 75% of all various types of items" stored at its fulfillment centers, "at speeds comparable to" front-line employees, and calls a human to "tag in" when it cannot handle an item. Its robotics overview describes Sparrow as an arm that moves individual items from containers into totes, and its Shreveport fulfillment center page says the latest version can handle more than 200 million unique products. Neither figure is a share of picks. Coverage of 75% of item types could mean more or less than 75% of daily picks, depending on which items sell most. Amazon gives no pick rate, no deployed count for either robot and no exception rate.

Counterevidence and limits

• Release dates can disagree. Otto's partnership release is dated 09.05.2023 on its site, while its January 2024 story says the partnership was announced in August 2023. This analysis uses "2023."
• Absence is not proof. No Otto statement giving a Covariant station count was found, but Otto may have published one in German or in channels not searched. Zalando's annual report is silent on the robots, but annual reports leave out many operating programs.
• The arithmetic is illustrative. The reconciliation of 10,000, 100,000 and 2 million a month rests on Nomagic's listed speeds, a 30-day month and an assumed reading of each release. It is not a finding about the Zalando pilot.
• Small, fashion-heavy sample. All three Nomagic customers named here are fashion companies. Rates in grocery, pharmacy or electronics may differ. Ocado's numbers are grocery-specific.
• Not a ranking. This piece does not compare vendors' performance, and the figures cited are not comparable across vendors because the units differ.
• No labor or ROI estimate. Nothing here is converted into labor savings, cost per pick or payback. Nomagic says it builds a "tailored ROI model" for prospects; that model is the vendor's.

What a buyer should ask for

A robot pick-rate claim becomes evidence when it carries its denominator. These are the fields that were missing from at least one figure in this analysis, and the evidence that would fill each:

Many warehouse robots sold for picking do not pick items themselves; they carry totes to people or bring shelves to stations. For what DHL, Locus and Geek+ customers confirm about those fleets, see our guide to collaborative picking robots.

The next checkpoints are Ocado's full-year 2026 results, Zalando's 2026 annual report and any Zalando update on the 50-robot roll-out. This guide will be reviewed when one of those is published, or within 90 days. For related coverage, see warehouse robot news and robot AI coverage.

How the sources were used

Customer releases and the customer's annual report establish what each customer said. Vendor releases and product pages are treated as company statements, including where they quote customers. Ocado and Amazon are treated as companies describing their own systems. Trade press is cited only for what it attributes to a named company. Calculations are simple divisions of published figures and are labeled as Servospace arithmetic. Corrections can be sent through Servospace's contact page.