Otto Group and robotics software company Sereact have announced a two-phase plan to bring NVIDIA-powered physical AI to one of retail’s hardest manual jobs: inspecting and sorting returned customer orders. Announced on October 28, 2026, the project will start with ten AI-controlled dual-arm systems at Otto Group’s logistics centre in Poland, with 15 more robotic arms planned in a second phase if the first results support it, taking the programme to 25 systems in total.
The two companies are developing and refining the technology together, and Otto Group will evaluate a broader rollout across its logistics network if the second phase succeeds. The announcement puts the focus on a part of e-commerce that shoppers rarely see but retailers feel every day: returns arrive folded, crumpled and repackaged, and the range of products never stops changing.
Key Takeaways
- Otto Group and Sereact will deploy ten AI-controlled dual-arm systems for returns processing in Poland, with 15 more planned in a second phase, for 25 in total.
- The systems use Sereact’s Cortex physical AI model, trained and validated with NVIDIA Isaac Sim, Isaac Lab and Cosmos, to handle items it has never seen before.
- Under a Robotics-as-a-Service model, Sereact operates and maintains the systems, while employees move to supervision, exception handling and quality checks.
Why Returns Are So Hard to Automate
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, and neither exists on a returns line.
That is the problem Sereact’s Cortex model is built to solve. Rather than relying on a fixed catalogue of items it already knows, Cortex is designed to generalise to objects it has never encountered. In real time, it decides how to grasp each item, assess its condition, and route it accordingly. For a returns operation, that combination matters: every item that arrives is, in a sense, a new case, and the system has to make a judgement about it on the spot.
The work is also a natural test case for physical AI, the branch of artificial intelligence that moves from screens and chatbots into machines that perceive and act in the physical world. A warehouse robot cannot simply recognise a product; it has to pick it up without damaging it, judge whether it is in saleable condition, and send it to the right next step. Getting any one of those steps wrong creates rework, waste or a disappointed second customer.
Few processes in e-commerce are as hard to automate as returns, because items come back folded, crumpled and repackaged, and the assortment never stops changing.
How the NVIDIA Technology Works
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. In other words, one real grasp can become many simulated examples, covering the messy variations that a returns line actually produces.
New policies are post-trained on NVIDIA Blackwell GPUs and tested against held-out cases before they reach the robots. That testing step is important: a policy that works on the examples it was trained on still has to prove itself on cases it has not seen. Only after that check does an update go live on the warehouse floor.
The feedback loop is designed to be fast. When a robot struggles with an item, the fix can come back as a tested update within days instead of weeks. That matters in retail, where seasons, promotions and product ranges change quickly, and a system that takes weeks to learn a new kind of return would always be behind the parcels arriving at the door.
- Phase one: ten AI-controlled dual-arm systems at Otto Group’s logistics centre in Poland.
- Phase two: 15 more robotic arms planned, based on initial results, for a total of 25 systems.
- Model: Sereact’s Cortex physical AI, which generalises to objects it has never encountered.
- Training stack: NVIDIA Isaac Sim, Isaac Lab and Cosmos, with post-training on NVIDIA Blackwell GPUs.
- Operating model: Robotics-as-a-Service, with Sereact operating and maintaining the systems.
- Human role: employees hand over repetitive grasping and lifting and take on supervision, exception handling and quality checks.
What Changes for Warehouse Workers
Under the Robotics-as-a-Service model, Sereact operates and maintains the systems. Employees hand over repetitive grasping and lifting and take on supervision, exception handling and quality checks. That division of labour is the project’s answer to the usual automation question: the robots take the most repetitive physical work, while people handle the cases that need judgement and the quality decisions that determine whether a returned item can be resold.
The same pattern is appearing across warehouse automation. Readers following the wider shift can see it in our earlier coverage of GXO and Agility Robotics expanding humanoid deployments, where robots are also being placed alongside people in logistics work rather than replacing the whole operation at once. The Otto and Sereact plan follows that phased logic: start with a defined task, measure the results, and only then scale.
Returns are a useful place to start because the cost of getting them wrong is visible. A returned item that is handled slowly ties up stock; one that is graded incorrectly either goes back on sale when it should not, or is discarded when it could have been resold. A system that can grasp, assess and route reliably therefore touches revenue, waste and customer trust at the same time.
Why This Fits the Wider AI Investment Wave
The announcement also lands in the middle of a much larger wave of AI infrastructure spending, in which NVIDIA’s hardware and software stack sits at the centre of many deals. The same week has seen major AI financing and partnership news, including NVIDIA’s $3 billion Meta AI partnership, showing how the company’s technology now reaches from data centres into warehouses and other physical workplaces.
For Otto Group, the test will be practical rather than theoretical. The first phase has to show that dual-arm systems can keep pace with a real returns line, handle the variety of items that arrive, and improve over days rather than weeks as new cases appear. If it does, the second phase and a broader rollout become a straightforward next step. If it does not, the phased structure limits the risk: ten systems in one logistics centre are a trial, not a company-wide replacement.
For the wider retail industry, the project is worth watching for the same reason. Returns processing is a shared headache, and a system that genuinely copes with unknown, crumpled and repackaged items would be useful far beyond one logistics centre in Poland. The question is whether physical AI has reached the point where it can handle that messiness day after day, at warehouse speed.
What to Watch Next
What to watch next is what this article has already set out: how the ten dual-arm systems perform at Otto Group’s logistics centre in Poland, whether the initial results support the planned second phase of 15 more arms for a total of 25, whether tested updates really return within days when a robot struggles with an item, and whether Otto Group then evaluates a broader rollout across its logistics network. Those results will show whether NVIDIA-powered physical AI has found one of its first large-scale jobs in retail, or whether returns remain too messy even for robots trained on thousands of simulated variants.