Ten first, fifteen if it works
The structure of the deal is the most informative part of it. Otto and Sereact announced a two-phase program covering up to 25 robots, not a firm deployment of all 25. The first phase calls for ten dual-arm systems. Another 15 are planned depending on the initial results, with any wider rollout across Otto's logistics network still under evaluation. Each step has to be earned by the previous one.
That caution fits the task. Returns processing is where warehouse automation usually goes to die. A fulfillment robot picks known products from known locations. A returns robot opens a parcel and finds a garment folded, crumpled or repackaged, in an assortment that changes every season. The partners say this plainly in their own release: traditional automation depends on known products and predictable presentation, and returns offer neither.
The proposed systems are intended to perform several connected steps in returns processing: grasping items, identifying them, assessing their condition and directing them for sorting. The release does not establish that the robots automate the entire returns workflow. Condition assessment is the step with judgment in it, and it is also the step where a mistake costs money, because an item graded wrong is either refunded wrongly or resold wrongly.
A catalog problem, answered with simulation
Sereact's answer, as described in the release, is a model called Cortex that is built to generalize. The company says it can recognize and handle objects it has never encountered, decides in real time how to grasp each one, and needs no product-specific teaching and no fixed item catalog. "A robot that only knows yesterday's catalog is of little use in a returns center," Sereact chief executive Ralf Gulde said in the release. "Cortex has to deal with whatever comes out of the next parcel."
The training method named in the announcement is simulation at volume. Real picks from the returns line are converted into thousands of simulated variants using NVIDIA Isaac Sim and Isaac Lab, then re-rendered with NVIDIA Cosmos world models across different garment states, packaging, materials and lighting. According to the partners, new policies are post-trained using NVIDIA Blackwell GPUs and tested against held-out cases before deployment. They say the process can turn a failure case into a tested update within days rather than weeks; no independent timing or reliability results were published.
Read those as vendor claims, because that is what they are. The release is the only source for how Cortex performs, and it offers no measurement of that performance: no items per hour, no grasp success rate, no accuracy figure for the condition grading. What is checkable today is the deployment structure, the site and the business model, not the capability.
Otto rents the robots, and keeps the judgment calls
Under the Robotics-as-a-Service arrangement in the release, Sereact operates and maintains the systems. Under the Robotics-as-a-Service arrangement, Sereact will operate and maintain the robotic systems. The announcement does not disclose the service fees, performance guarantees or how operational responsibilities are divided beyond that description. Sereact says its systems use experience from returns handling to improve subsequent performance. The announcement does not explain whether updates developed at Otto's site can be reused for other customers, or how customer data is separated and governed. The release's About section lists Sereact's customers as including Zalando, IKEA and DHL.
The partners say the intended arrangement would shift employees away from repetitive grasping and lifting toward supervision, exception handling and quality checks, with final quality decisions remaining with people. The release gives no staffing projections or measured employment effects. Otto frames the project as an experiment at working scale. "This project lets us find out, at real scale, how far Physical AI can take us," supply chain executive Martin Umland said in the release. It sits inside a group-wide commitment to invest €350 million in technology and AI, which is the context for the spending, not this project's budget.
Sources
- [1] Otto Group and Sereact, "Otto Group and Sereact Bring NVIDIA-Powered Physical AI to Returns Processing," GlobeNewswire, October 28, 2026 (phase structure and counts, Poland site, Cortex description, NVIDIA training stack, update-loop claim, Robotics-as-a-Service model, staffing description, both executive quotes, €350 million group commitment, Otto revenue, Sereact Series B and customer list; the partners' own announcement, read in full first-hand this run)Read source