Research · October 28, 2026

Otto × Sereact: Physical AI Takes On the Returns Line

On October 28, 2026, Otto Group and Sereact announced that up to 25 AI-controlled dual-arm robots will take over the inspection and sorting of returned orders at an Otto Group logistics center in Poland. Ten systems go in first. Fifteen more follow in a second phase, but only if the first phase earns it — and a broader rollout across Otto's logistics network is on the table only if the second phase succeeds. The structure of the announcement is the story: this is a staged, gated deployment, not a purchase order.

Returns break traditional automation. Few processes in e-commerce are as hard to automate as returns. Items come back folded, crumpled, and repackaged, and a multi-category retailer's assortment never stops changing. Traditional automation depends on known products and predictable presentation. Returns offer neither. That is why the returns line, for all the money spent on warehouse automation, has stayed stubbornly manual — and why it is a revealing place to test whether physical AI is real or a demo.

The claim is generalization, not a bigger catalog. The robots run on Cortex, Sereact's Physical AI model, which is built to recognize and handle items it has never encountered — no product-specific teaching, no fixed item catalog. Each unit decides in real time how to grasp an item, assess its condition, and route it. Sereact CEO Ralf Gulde put the point sharply: "A robot that only knows yesterday's catalog is of little use in a returns center." And then the sentence that carries the whole investment case: "Every return the systems handle makes the model better at the next one." If that flywheel is real, the moat is not the arm. It is the accumulated experience of handling the parcels nobody could pre-program.

The update loop runs at software speed. Cortex is built on NVIDIA's physical AI stack. Sereact uses NVIDIA Isaac Sim and NVIDIA Isaac Lab to turn real picks from the returns line into thousands of simulated variants, and NVIDIA Cosmos world foundation models to re-render them across garment states, packaging, materials, and lighting. New policies are post-trained on NVIDIA Blackwell GPUs and tested against held-out cases before they reach the robots. When a robot struggles with an item, the fix can come back as a tested update within days instead of weeks. The warehouse moves at conveyor speed; the model improves at software speed. That gap is the speed advantage, if it holds at Otto's volumes.

The business model removes the adoption risk. Under a Robotics-as-a-Service model, Sereact operates and maintains the systems. Otto does not buy a fleet and hope. Employees shift away from repetitive grasping and lifting toward supervision, exception handling, and final quality decisions — the tasks that depend on experience and judgment. The context behind the deal is Otto's medium-term commitment to invest 350 million euros in technology and AI. Otto Group, a family-owned German retail group with more than 30,000 employees and 13.8 billion euros in revenue in its 2025/26 financial year, is not a startup's pilot customer. It is a scaled operator buying an option on a capability, in stages, with exit ramps at every gate.

Otto is the second proof point, not the first. Three weeks earlier, on October 7, 2026, CEVA Logistics announced that Sereact's dual-arm systems had entered live operations handling Zalando fashion returns at sites in Greven, Germany, and Swiebodzin, Poland — delivered under the same Robotics-as-a-Service model, with the same no-per-item-training claim, described as the first phase of a broader initiative to scale returns automation across Europe. Otto is therefore not a pilot in a vacuum. It is the second deployment of the same system, with a different customer, in a different network. The pattern to watch is whether phase-gated rollouts compound: CEVA to Otto, Otto's first phase to its second, and then, if the results hold, to a network.

What this is not. It is not a measured result. No throughput, accuracy, cost-per-return, labor-saving, or price figures were disclosed with the announcement, and none are claimed here. There is no date for the second phase and no commitment to the broader rollout — both are conditional on results the partners have not yet published. The capability claims are the vendors' own; nothing here has been independently measured. The test that matters is narrow and observable: does phase one convert into phase two, and does handling quality at Otto's volumes confirm the flywheel Gulde describes? Until then, this is a well-structured bet on physical AI, placed by a serious operator, at the hardest workstation in the warehouse.

Neither Otto Group nor Sereact is scored by BlitzLab — Otto is a private German retail group, Sereact a private robotics startup — so this article carries no score and changes none. NVIDIA, whose Isaac, Cosmos, and Blackwell technologies form the stack, is mentioned only as the technology provider; its posted score is unchanged.

Sources