Otto and Sereact Put NVIDIA Physical AI Robots on Returns Sorting
Key Takeaways
- Otto Sereact Physical AI robots will take on one of the hardest jobs in e-commerce: inspecting and sorting returned customer orders at Otto Group’s logistics centre in Poland.
- The rollout is planned in two phases: 10 AI-controlled dual-arm systems first, then 15 more based on initial results, for a total of 25.
- The robots run Sereact’s Cortex model, a Physical AI system built on NVIDIA’s stack, including Isaac Sim, Isaac Lab, and Cosmos world foundation models.
- Returns defeat traditional automation because items arrive folded, crumpled, and repackaged, with an assortment that never stops changing.
- Sereact will operate and maintain the systems under a Robotics-as-a-Service model, while warehouse staff shift to supervision, exception handling, and quality checks.
- Otto Group will evaluate a broader rollout across its logistics network if the second phase succeeds.
Otto Group and robotics software company Sereact announced on 28 October 2026 that they will deploy AI-controlled robots for returns processing at an Otto Group logistics centre in Poland. The announcement matters because returns are one of the few warehouse jobs that have resisted automation for years. Everything about a returned parcel is unpredictable, and that unpredictability is exactly what the new systems are designed to handle.
Most warehouse robots work best when they know what they are about to pick. A product has a fixed shape, a fixed barcode position, and a fixed storage slot. Returns break every one of those assumptions. A dress comes back folded differently from how it left. A shoe box comes back dented, taped shut a second time, or missing its lid. A multi-category retailer adds new products constantly, so the range of items passing through the returns line changes week by week. Traditional automation depends on known products and predictable presentation, and returns offer neither.
That is the gap Sereact’s system is built to close. Its Physical AI model, called Cortex, is designed to generalize to objects it has never encountered. When a robot picks up a returned item, the model decides in real time how to grasp it, assesses its condition, and routes it to the right next step. There is no per-item training run and no fixed profile for each product. The robot looks at the item in front of it and works out what to do, much as a human worker would, but at machine speed and without fatigue.
Why Returns Are the Hardest Job in the Warehouse
It helps to picture what actually arrives at a returns station. Outbound picking is orderly. A worker or robot takes a known item from a known location and puts it in a box. Returns run the process in reverse, minus the order. Items arrive in bags, in boxes, in their original packaging, or in no packaging at all. Some are in perfect condition and can go straight back to stock. Others are worn, damaged, or incomplete and need a different route. Some need cleaning or repair before they can be sold again.
Each of those decisions requires judgment. Is this garment resalable as new, or should it be marked down? Is this item complete, or is a part missing? Is the packaging good enough to ship again? Human workers answer these questions by looking, touching, and comparing. Reproducing that judgment in a machine is a far harder problem than moving a box from one conveyor to another, and it is why returns departments have stayed labour-intensive even as the rest of the warehouse automated.
There is also a regulatory reason to get returns right. In the European Union, rules restrict the destruction and landfill disposal of returned merchandise, which pushes retailers to recondition returned goods for resale wherever possible. A system that can inspect, grade, and route returns quickly is therefore not just a labour saving. It helps retailers recover value from stock that might otherwise be written off.
What Otto and Sereact Are Actually Deploying
The plan has two phases. In the first phase, 10 AI-controlled dual-arm systems will be deployed at Otto Group’s logistics centre in Poland. In the second phase, 15 more robotic arms are planned to be added, based on the results of the first phase, bringing the total to 25. The two companies are jointly developing and refining the technology as part of the project, and Otto Group is evaluating a broader rollout across its logistics network if the second phase succeeds.
Two details in that structure are worth noting. First, the second phase is conditional, not guaranteed. The extra 15 arms depend on what the first 10 prove in real operation. That is a sensible way to buy robotics: prove the system on one line, then scale it. Second, the deployment is a joint development project, not a simple equipment purchase. Otto Group is not just buying robots off a shelf. It is working with Sereact to refine the technology against its own returns flow, which means the system should get better at Otto’s specific mix of products over time.
The commercial model matters too. Under a Robotics-as-a-Service arrangement, Sereact operates and maintains the systems. Otto Group pays for robotic capacity rather than owning and servicing the machines itself. That shifts the maintenance burden to the company that built the software, and it means upgrades to the AI model flow to the robots as part of the service rather than as a separate project.
This is the same pattern already running at other European logistics sites. CEVA Logistics announced on 7 October 2026 that Sereact dual-arm systems had entered live operations at sites in Greven, Germany, and Świebodzin, Poland, handling returns for fashion platform Zalando. Those systems grasp, identify, and sort fashion returns without per-item training or fixed product profiles, and Zalando joined Sereact’s Series B funding round as a strategic investor in July 2026. The Otto project extends a model that is already working in live logistics centres rather than starting from a laboratory demo.
How Cortex Learns Without Knowing the Product
The technical core of the deployment is the training loop. Sereact uses NVIDIA Isaac Sim and NVIDIA Isaac Lab to turn real picks from the returns line into thousands of simulated variants. NVIDIA Cosmos world foundation models then re-render those variants across different garment states, packaging types, materials, and lighting conditions. New policies are post-trained on NVIDIA Blackwell GPUs and tested against held-out cases before they reach the robots.
In plain terms, the system learns from reality and then practises in simulation. A robot’s real attempt at a difficult item becomes training data. Simulation multiplies that one attempt into thousands of variations: the same item folded differently, lit differently, packaged differently. The updated model is tested on cases it has not seen before it is allowed near a live returns line. When a robot struggles with an item, the fix can come back as a tested update within days instead of weeks.
That speed is the real breakthrough. Older warehouse robots improved slowly because every new product needed explicit programming or a dedicated training run. A model that generalizes to unseen objects, and that can be updated in days from real operating data, changes the economics of automating messy, variable work. The returns line is the proving ground because it is the messiest, most variable work a warehouse has.
For readers who follow autonomous machines more broadly, this is the same direction as other physical AI deployments we have covered, including driverless heavy equipment. Our article on Bedrock Robotics and its $600M round for autonomous excavators looks at machines that work without operators on live sites, while our piece on AiMOGA Mobo, the humanoid companion robot shows the consumer end of the same robotics wave. The Otto deployment sits between those extremes: industrial work, but work that requires judgment, not just movement.
What Changes for Warehouse Workers
Automation announcements usually raise the same question, and the companies have answered it directly. Employees hand over repetitive grasping and lifting and take on supervision, exception handling, and quality checks. The robots do the reaching, lifting, and sorting. People watch the line, deal with the items the robots cannot resolve, and check quality.
That division of labour makes sense for returns. Even a good model will meet items it cannot classify confidently: a garment with an unusual defect, a product with missing parts, a package with no readable label. Those edge cases go to a human, and the human’s decision becomes more training data. Over time the share of items needing human help should fall, but the human role does not disappear. It moves up the skill ladder, from handling every item to handling the exceptions.
It is also worth being honest about what the announcement does not say. It does not publish throughput figures, such as items handled per hour, and it does not publish cost figures or a payback period. Those numbers will only emerge from the first phase of operation. Anyone evaluating this deployment, whether as a retailer or as an observer of the robotics market, should watch for them. The claim that matters most, that the system generalizes to unseen items at production speed, can only be proven by sustained operation, not by a press release.
What to Watch Next
Three milestones will tell us whether this deployment is a genuine step forward. The first is the completion of phase one: 10 systems running day to day on Otto’s returns line in Poland. The second is the phase two decision: whether the initial results justify adding 15 more arms to reach 25. The third is the broader rollout question: whether Otto Group extends the technology beyond the Poland centre to the rest of its logistics network.
The companies also demonstrated the technology at NVIDIA GTC Berlin 2026, where Cortex ran on two Universal Robots arms performing a returns-handling application already implemented in live logistics centres. Trade-show demos are easy to dismiss, but this one is notable because the application shown is the same one running in production for Zalando returns. The gap between demo and deployment is smaller here than it usually is.
The bigger picture is that physical AI is moving from vehicles and humanoids into the unglamorous middle of the economy: warehouses, returns departments, and logistics centres. Those are the places where labour is hardest to hire and where the work is hardest to automate. If a model can learn to handle a crumpled, repackaged, unknown item on a returns line, it can handle a large share of the physical work that still depends on human hands. Otto’s Poland deployment is a test of that idea at real scale, and its results will be watched well beyond retail.
