Loading page…
The partners say retraining can shrink from weeks to days. Simulated checks come first, but each passing update still faces a test on the returns line.
Listen to this story
The October 21 launch connects recorded robot picks to simulated data generation, Cortex retraining and held-out evaluation, with passing policies advancing to station tests at selected CEVA fulfillment centers handling Zalando returns. Sereact says the workflow could cut an earlier weeks-long, manually staged retraining cycle to days, but passing simulation earns a real-world test—not proof of success. Nebius is offering the workflow as a reusable Workbench reference architecture on its EU cloud; engineers set the goals and approve runs above configured GPU-spending limits.
Isaac Lab produces thousands of simulated pick variants, while Cosmos varies garment appearance, packaging, materials and lighting.
The retrained policy runs on NVIDIA Blackwell GPUs and must pass Sereact’s threshold on simulated cases withheld from training.
After a failed evaluation, the agent can generate targeted examples for missed cases and repeat training; engineers still define the validation goal.
A mistake by a robot handling Zalando returns could become a tested software update within days rather than weeks, Sereact says. That is the promise behind a training workflow Sereact, NVIDIA, Nebius, CEVA Logistics and Zalando announced on October 21, initially targeting returns handling at selected CEVA fulfillment centers.
Returned clothes do not arrive in a standard condition. They may be folded or crumpled, inside a plastic bag or loose, with tags attached or missing. A robot must pick up an item, identify it, inspect its condition and sort it. The assortment also changes with the season.
Sereact supplies the dual-arm returns stations and Cortex, its robot-control model. Cortex combines vision, language and action, and learns from recorded picks. But those recordings cover only items that have already reached a station. In the previous process, engineers collected examples on site, retrained the model, tested it on a robot and repeated the cycle.
The joint announcement describes that earlier cycle as taking weeks and reaching only the cases a team could stage by hand. The new workflow connects recorded robot activity, simulation, further training and evaluation through one shared data layer, rather than treating each as a separate step.
NVIDIA’s Isaac Lab turns the recorded episodes into thousands of simulated variants. Cosmos then changes their visual presentation across garment conditions, packaging, materials and lighting. This gives the training process variations beyond the particular item and conditions captured during the original pick.
The resulting dataset is versioned and used to further train Cortex on NVIDIA Blackwell GPUs. The updated policy—the model’s instructions for how the robot should act—is then scored against simulated variants held out of training. Sereact sets the threshold it must pass.
Passing that threshold earns an update a real-world test at the returns stations, not an automatic declaration of success. A failed evaluation instead sends the workflow back to generate targeted examples of the cases it missed and train again. Zalando’s returns operation is the setting where the updated behavior must ultimately prove itself.
An AI agent builds and operates the workflow using NVIDIA’s Physical AI Data Factory skills: open, versioned instruction files a coding agent can read. Nebius adds instructions for simulation, training, evaluation and checking whether an update passes. The agent launches stages, checks results and chooses which cases need more generated examples after a failed run.
Engineers still specify the goal, such as validating behavior for an autumn assortment. The agent previews GPU costs before each run, and engineers must approve any run above a set spending limit. That separates the agent’s job of organizing the work from human decisions about its purpose and larger computing bills.
Nebius supplies its open-source Physical AI Workbench, shared storage and computing infrastructure through EU data centers. The pipeline is available from October 21 as a Workbench reference architecture—a prepared design other robotics teams can use—on Nebius AI cloud, with NVIDIA’s instruction files pre-installed. Customers pay for the computing and storage they use; access is not limited to the Zalando deployment.
Loading discussion...
Join the conversation
Explain what reassurance you would want before a trial begins.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.