Industries / Transport and vehicles / Warehousing and distribution
Returns
Returns work takes back what customers send back, inspects and grades each item, and decides whether to restock, repair, resell or dispose of it.
5
Milestones achieved, from 2019 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2019
Laundry-folding robot maker fails
Seven Dreamers, the Japanese company behind Laundroid, a home laundry-folding machine shown at trade shows for years, filed for bankruptcy in Japan with roughly $22 million in debt, by TechCrunch's figure. The machine never shipped, and hands-on reports said it struggled with basic tasks.
LessonFolding soft, floppy garments is much harder for a machine than it looks, which matters for any returns center that wants robots to refold clothes.
Credit
Seven Dreamers
2022
France bans destroying unsold goods
France's anti-waste law of 10 February 2020, known as AGEC, ended the burning and landfilling of unsold non-food goods; they must be reused first, or recycled if reuse fails. The ban applied from 1 January 2022 to products covered by producer take-back schemes, hygiene and baby-care products, books and some other groups, and from 31 December 2023 to everything else. France's environment ministry calls the measure "a world first."
Credit
French government
2023
UPS buys box-free returns
UPS bought Happy Returns, founded in Los Angeles in 2015, from PayPal. Shoppers hand items over at a counter with no box or label, and the returns are pooled, sorted in automated buildings and shipped back to each retailer in bulk. In its 2025 survey with the National Retail Federation, Happy Returns reported that retailers expected 15.8% of their sales, $849.9 billion, and 19.3% of online sales to come back that year; those are the retailers' own estimates.
Credit
Happy Returns; UPS
2026
EU bans destroying unsold clothes
From 19 July 2026 large companies in the European Union may not destroy unsold apparel, clothing accessories or footwear; medium-sized companies follow from 19 July 2030, and micro and small companies are exempt. "Unsold" includes goods that consumers returned under their right to withdraw from a purchase. Recycling and burning count as destruction, while refurbishing does not, and companies must publish the number and weight of unsold goods they discard each year.
Credit
European Parliament and Council, Regulation (EU) 2024/1781 (Ecodesign for Sustainable Products Regulation), Article 25
2026
Robots sort fashion returns
Zalando and CEVA Logistics said two-armed robots from Sereact had started live work on returned fashion items at CEVA sites in Greven, Germany, and Świebodzin, Poland. The robots identify, grasp and sort items they have not been trained on, which matters because returns arrive crumpled and repacked. No speed or accuracy figures were published.
Credit
Sereact; CEVA Logistics and Zalando as customers
Watch for
Robots inspect returns at Otto
Otto Group will install 10 of Sereact's two-armed robots to inspect and sort returns at its logistics center in Poland, then add 15 more if the first results are good, and will consider a wider rollout after that.
Achieved whenOtto publishes results from the second phase, such as items handled per hour and the share graded as a human inspector would, and decides to roll out more widely.
Claimed by
Otto Group and Sereact, November 2026
AI against return fraud
In a survey of 358 e-commerce staff at large US merchants, 85% said they use artificial intelligence to detect or prevent return fraud, and the same report estimates that 9% of returns are fraudulent.
Achieved whenA retailer or an independent auditor publishes measured results for an AI screening system in a returns center, giving both the fraud it caught and the honest returns it wrongly flagged.
Claimed by
Retailers surveyed by the National Retail Federation and Happy Returns, October 2025
Open
01
A graded set of returns
An open image set of returned or open-box household goods (not clothing). Each item is photographed from fixed angles on a measuring grid, and two people separately grade it into a resale grade (like new, open box, used, damaged or missing parts) and a next step (restock, refurbish, resell as is, recycle), following a one-page written rule. The project publishes how often the two graders agree, then tests a vision-language model against the grades both graders gave.
The briefCredit · unclaimed02
Fold more than T-shirts
A two-armed, low-cost robot that takes a crumpled returned garment from a bin and folds it to a shop's standard fold, tested on four garment types: T-shirts, long-sleeved shirts, trousers and hooded sweatshirts, with every attempt filmed and tallied.
The briefCredit · unclaimed
Open 01 · the brief
A graded set of returns
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open image set of returned or open-box household goods (not clothing). Each item is photographed from fixed angles on a measuring grid, and two people separately grade it into a resale grade (like new, open box, used, damaged or missing parts) and a next step (restock, refurbish, resell as is, recycle), following a one-page written rule. The project publishes how often the two graders agree, then tests a vision-language model against the grades both graders gave.
Why it is still open
The closest open datasets cover something else. Kaputt (2025) has over 230,000 images of retail logistics items labeled defective or not, without resale grades. The Second-Hand Fashion dataset from RISE and Wargön Innovation covers 31,997 donated garments with a 1-to-5 condition score. Neither covers boxed e-commerce returns.
Built with what exists
A phone camera; a printed grid mat; the written rule; a spreadsheet; a vision-language model (see the Note "Three working devices"); a reviewer model that flags doubtful grades for a person to check (see the Note "Seven more flagged").
First step
Write the one-page rule, and have two people grade the same 50 items separately to see whether the rule produces the same grades.
Done looks like
500 items, each graded by two people, released under an open license, with the agreement between graders (Cohen's kappa, a standard agreement score) and the model's accuracy against the grades both graders gave.
Open 02 · the brief
Fold more than T-shirts
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A two-armed, low-cost robot that takes a crumpled returned garment from a bin and folds it to a shop's standard fold, tested on four garment types: T-shirts, long-sleeved shirts, trousers and hooded sweatshirts, with every attempt filmed and tallied.
Why it is still open
A Hugging Face team reported in 2026 that its two-armed robot folds a random T-shirt 90% of the time, over 20 filmed attempts in its own test, and released its code, data and models; it tested T-shirts only. In lab work, SpeedFolding (2022) folded garments from a random start with 93% success at 30 to 40 folds an hour. The Sereact robots at Zalando sort returns but are not reported to fold them, and Sewts, whose robot folds towels in laundries, names clothing returns only as a future plan.
Built with what exists
The Hugging Face team's released folding model and data as a starting point; the open-source OpenArm the team used, or two SO-101 arms with LeRobot (see the Note "No download"); a robot model that learns from a few demonstrations (see the Note "Five demonstrations per task") or open robot model weights (see the Note "Open robot model weights"); a folding-board template; an overhead camera.
First step
Rerun the Hugging Face T-shirt model on your own arms for ten attempts, then record 50 demonstrations of folding one long-sleeved shirt.
Done looks like
200 filmed attempts, ten on each of five garments of each of the four types, with a published tally of folds within 2 cm of the template and the time per fold, set beside the Hugging Face T-shirt result.