Fashion returns are difficult to automate for a simple reason: the item coming back rarely looks like the item that left the warehouse.

A shirt may be folded badly. A shoe can arrive without its original packing. Different sizes, fabrics and shapes appear in no predictable order.

That has made returns one of the harder warehouse jobs to hand over to conventional robots.

Zalando, CEVA Logistics and Stuttgart-based robotics company Sereact are now testing a different approach at production scale.

On 7 October, the companies announced that dual-arm robotic systems had entered live returns operations at CEVA sites handling Zalando returns in Greven, Germany, and Świebodzin, Poland.

The robots use Sereact’s Cortex software to identify, grasp and sort returned fashion items without needing a fixed SKU list or article-specific training.

That is the interesting part.

The companies have not published independent figures showing faster returns, lower cost per item or fewer errors.

So the useful question is not whether the robots look impressive.

It is whether they make the returns operation measurably better.

Why fashion returns are harder than outbound fulfilment

Outbound fulfilment starts with more certainty.

The warehouse generally knows what the item is, where it is stored, what packaging it needs and where it should go.

Returns start with uncertainty.

An item may arrive:

  • folded differently;
  • partly unpacked;
  • inside the wrong packaging;
  • mixed with other items;
  • facing an unexpected direction;
  • damaged;
  • missing a tag;
  • requiring a quality check.

Traditional warehouse automation works well when the environment is predictable.

Returns often are not.

CEVA describes returns as particularly difficult because of unpredictable volumes, manual quality checks, seasonality and high cost per unit.

Those are company claims, but the underlying operational problem is easy to understand.

The same robot cannot simply repeat one fixed movement thousands of times if every parcel contains something different.

What the robots are doing

Sereact’s setup uses two robotic arms powered by its Cortex software.

According to the companies, the system can deal with objects it has not seen before.

It identifies the item, works out how to grasp it and then sorts it without requiring engineers to build a separate handling routine for each product.

That removes one of the obvious barriers to fashion automation.

A retailer with thousands of SKUs cannot realistically train a robot one product at a time, particularly when the returned item may no longer be packaged as expected.

The systems are now operating at CEVA facilities in Germany and Poland rather than remaining in a lab demonstration.

That gives retailers something more useful to watch.

The question shifts from “Can a robot pick up clothing?” to “Can it do useful work repeatedly inside a real returns process?”

Do not mistake live operation for proven economics

There is an important gap in the announcement.

The companies do not publish figures for:

  • items processed per hour;
  • successful handling rate;
  • exception rate;
  • cost per processed return;
  • error rate;
  • damage rate;
  • time from receipt to resale;
  • percentage of the returns flow handled automatically.

Without those numbers, there is no basis for saying the deployment has made Zalando’s returns cheaper or faster.

The announcement confirms that the systems are operating.

It does not prove the business case.

That distinction should remain clear as more warehouse robotics stories appear.

A useful automation system has to do more than perform the task.

It has to perform enough of the task, reliably enough, at a cost that makes sense.

Measure the exception rate first

For retailers looking at similar systems, one of the most useful numbers may be the exception rate.

That is the share of items the robot cannot process without human help.

Imagine two systems.

System A processes 600 items per hour but sends 25 percent of them to a person.

System B processes 450 per hour but sends only 5 percent to a person.

The faster robot is not automatically the better operation.

The exception queue still needs people, space and time.

For fashion returns, log why each item leaves the automated flow.

Possible reasons include:

  • damaged item;
  • missing identifier;
  • unusual shape;
  • uncertain classification;
  • failed grip;
  • packaging obstruction;
  • quality review needed;
  • system confidence too low.

That data tells a retailer whether the difficult cases are rare edge cases or a large part of everyday returns.

Track whether items get back to sale faster

The commercial goal of returns automation is not simply moving clothing between bins.

For many retailers, time matters because a returned item may still be sellable.

The longer it sits in the returns process, the longer that inventory is unavailable.

A useful metric is therefore:

Return-to-resale time

Measure the time from the return entering the facility until a sellable item is available in inventory again.

Then compare:

  • robot-handled items;
  • manually handled items;
  • items requiring exceptions.

If automation reduces handling time at one station but creates a queue at quality control, the end-to-end result may barely change.

The whole process should be measured, not one robotic movement.

Measure accuracy, not just speed

Returns operations make decisions that can affect inventory quality.

A system that moves items quickly but sorts them incorrectly can create more work later.

Useful checks include:

Metric What it tells you
Successful grasp rate Whether the robot can physically handle the mix
Correct sort rate Whether items reach the right next step
Exception rate How often people must intervene
Reprocessing rate How often work has to be repeated
Damage rate Whether automated handling harms goods
Return-to-resale time Whether sellable stock comes back faster
Cost per return Whether the economics improve
Human touches per return Whether repetitive handling actually falls

Do not evaluate these once.

Returns volumes and product mixes change with seasons, promotions and customer behaviour.

A system that performs well during an ordinary week may behave differently after Christmas or a major sales event.

The labour story needs measurement too

CEVA and Sereact say the robots take over repetitive handling while employees move toward station supervision, exception management and quality control.

That is plausible, but it should be measured rather than assumed.

Track what actually changes for workers.

Questions include:

  • How many repetitive lifts disappear?
  • How much time moves into exception handling?
  • Does the job require new training?
  • How many stations can one employee supervise?
  • Does physical strain fall?
  • Does troubleshooting create new workload?
  • What happens during peak returns periods?

Replacing repetitive movement with more skilled work can be valuable.

It is still a change in work design, not simply a technology upgrade.

Retailers should measure both sides.

The Robotics-as-a-Service model changes the buying decision

The deployment is being delivered through a Robotics-as-a-Service, or RaaS, model.

That can reduce the need for a retailer or logistics provider to buy the full system outright.

Instead, the commercial decision can look more like an ongoing operating cost.

For retailers assessing this model, compare more than the monthly robotics charge.

Include:

  • installation;
  • integration;
  • warehouse changes;
  • support;
  • downtime;
  • human supervision;
  • exception handling;
  • maintenance;
  • software updates;
  • peak-capacity requirements.

A lower upfront cost does not automatically mean lower total cost.

The advantage of RaaS is that it can make a live trial easier to start and expand if the numbers work.

Zalando already has a deeper relationship with Sereact

This is not a one-off warehouse experiment between unrelated companies.

Zalando joined Sereact’s Series B round as a strategic investor in July 2026.

The companies say the Germany and Poland deployment is the first operational stage of a broader effort to expand returns automation across their European logistics footprint.

That makes Greven and Świebodzin useful sites to watch.

If the economics work, similar systems could move into more facilities.

If they do not, the current sites should expose where the difficult parts remain.

The public evidence is not strong enough yet to say which outcome is more likely.

What ecommerce teams should ask before copying the model

A fashion retailer does not need to be Zalando to learn from the deployment.

Before buying similar technology, answer these questions.

How variable are our returns?

Look at product type, packaging, condition and seasonal variation.

Where is the current bottleneck?

If quality inspection is the slowest stage, automating item handling may not solve the main problem.

How many items need human judgment?

Separate physical handling from decisions about damage, resale condition or fraud.

What is our current cost per return?

Without a baseline, a robotics supplier cannot prove improvement.

How quickly does sellable stock return to inventory?

This often matters more commercially than the speed of one workstation.

What happens when the system fails?

Measure exception handling, downtime and recovery.

Can the technology handle our peak?

Average daily volume is not enough for a returns operation shaped by holidays and promotions.

This is a live deployment, not proof that fashion returns are solved

The Germany and Poland sites are significant because the robots are working in real operations with Zalando return volumes.

That is a stronger signal than a controlled demonstration.

It is still too early to call the problem solved.

The companies have supplied useful information about how the system works, but not enough performance data to judge the economics independently.

The next useful evidence will be operational.

How many items can the systems handle?

How often do people step in?

Does sellable inventory return faster?

Does cost per return fall after integration and support costs are included?

Those numbers will tell us whether physical AI is changing fashion returns or simply moving robotics into a harder part of the warehouse.

For more European marketplace and logistics developments, follow NEMO’s Ecommerce coverage.

NEMO has also covered Allegro’s 27 and 28 October seller changes and the emerging Personal Agent Protocol for merchants.

Primary sources