Sereact, NVIDIA, Nebius, CEVA Logistics and Zalando build closed-loop robot training for returns processing

Sereact, NVIDIA, Nebius, CEVA Logistics and Zalando announced on October 21, 2026, that they had built a closed training loop for robotic returns processing. The initial deployment will cover returns processing for Zalando at selected CEVA Logistics fulfillment centers. NVIDIA will provide capabilities from Isaac Sim, Isaac Lab, Cosmos and Physical AI Data Factory, while Nebius will operate Physical AI Workbench through EU data centers.

Secondary reports say that returns operations for Zalando and CEVA Logistics are among the more variable processes in e-commerce logistics, making their success a measure of the practical value of the robot training platform. The division of roles, with NVIDIA and Nebius providing computing and data infrastructure and Sereact providing the robot’s “brain,” shows that the success of a deployment depends not only on the robot itself but also on the surrounding infrastructure.

What is new in this report is that the companies have explicitly defined their respective roles in a closed loop that runs end to end from simulation and training to deployment in the field, targeting returns processing, a real-world process with substantial variability. At the same time, using NVIDIA’s Isaac Sim, Isaac Lab and Cosmos for the robot training platform, with Nebius operating Physical AI Workbench on EU data centers, follows the division of labor expected in physical AI development. The novelty appears to lie less in the technology itself than in applying this division of labor to returns processing, a logistics process with many exceptions. An earlier article from this newspaper, Teradyne Robotics and Sereact partner to expand physical AI for retail and logistics, reported on the partnership. This time, the collaboration is being applied to returns operations for Zalando and CEVA Logistics, making the use case more specific. In other words, Sereact’s external expansion is moving from the broad category of “logistics” to deployment in a specific process: returns handling. However, this announcement alone does not reveal how much data each company will share within the closed training loop or what level of automation they are targeting. At an industry level, the announcement shows that the competitive unit is becoming an infrastructure stack that connects not only robot performance but also simulation, world models, data centers and field operations. Returns processing is considered one of the more variable areas of logistics, so a key question is whether models trained on this process can be extended to other warehouse tasks. Implementation will depend on what data is stored in the Nebius environment, how well Cortex training can accommodate differences between sites, and how broadly the system will be applied to Zalando’s returns operations. Areas to watch include the number of robots deployed, the number of operating sites, the automation rate, how training data is handled, and plans to expand beyond returns processing.

For companies in Japan working on similar processes, this is an example of the need to coordinate not only robots but also simulation, world models and data center operations. How warehouse automation and returns-processing data is connected to the training platform is likely to be a key deployment issue.