CEVA and Zalando Cut Robot Training Cycles from Weeks to Days

CEVA Logistics and Zalando say a new AI training pipeline shortens robotic training and improvement cycles from weeks to days — a claim that, if it holds at scale, would reshape the economics of automating e-commerce returns handling.

Announced on October 22, 2026, the initiative builds a closed-loop training process for Cortex, the robotic AI system of German startup Sereact, combining real warehouse data with simulation, model training and performance evaluation. NVIDIA and cloud provider Nebius supply the simulation and computing layers.

The first deployment targets Zalando's returns handling at selected CEVA fulfillment centers, where robots process products arriving in varying conditions, packaging formats and configurations. Returns handling has long resisted automation precisely because items can arrive folded, crumpled, unpackaged or without original labels — a degree of unpredictability that traditional fixed-logic systems cannot absorb.

How does the training loop work?

The architecture pairs Sereact's Cortex vision-language-action model with NVIDIA's Isaac Sim, Isaac Lab and Cosmos technologies, running on Nebius' Physical AI Workbench.

The sequence runs as follows:

• Robotic stations at CEVA sites collect data from live returns handling — video recordings and robot movement telemetry.
• Datasets transfer to Nebius cloud infrastructure, where NVIDIA Isaac Lab generates simulated variations of real handling scenarios.
• NVIDIA Cosmos expands those datasets further, covering different garment conditions, packaging materials and lighting environments.
• Retrained on NVIDIA Blackwell GPUs, Cortex encounters a far wider range of handling situations before updated software ever touches the warehouse floor.
• Each updated policy is scored against predefined performance criteria set by Sereact. Passing versions go to real-world testing; failing versions trigger additional simulation and training targeted at the identified weaknesses.

An AI agent coordinates the pipeline using NVIDIA's Physical AI Data Factory skills and Nebius' Workbench framework. Humans remain in the loop: engineers set training objectives and must approve any operation exceeding predefined computing cost limits — a cost brake worth noting, since simulation-heavy training can escalate GPU bills quickly.

Who does what?

The division of labor across the four parties is precise. Sereact provides the Cortex AI model and dual-arm robotic returns stations. NVIDIA supplies simulation and training technologies. Nebius contributes cloud infrastructure and the development environment through its European data centers. CEVA Logistics hosts the robotic stations within its fulfillment operations for Zalando, and Zalando supplies the returns-handling use case against which the technology is evaluated.

What does it mean for 3PLs and retailers?

For CEVA, the commercial logic is straightforward: returns handling is labor-intensive, high-variability work that sits uneasily with conventional automation, and any credible path to robotizing it would cut cost per return while lifting throughput consistency. For Zalando, faster training cycles mean robotic systems can adapt more quickly to changing product assortments and seasonal demand — a chronic pain point in fashion e-commerce, where the SKU base turns over continuously.

The promise of compressing improvement cycles from weeks to days rests, for now, on the companies' own account. No throughput figures, deployment counts or returns-per-hour data accompanied the announcement, so any operational gains remain to be demonstrated in production. The parties say they aim to create a continuous improvement process in which data collected during live operations feeds the development and testing of updated robotic capabilities — the standard closed-loop pitch, but here grounded in a named use case rather than a lab demo.

A reference architecture for others

Beyond the Zalando deployment, the training pipeline is available as a reference architecture through Nebius' Physical AI Workbench. The framework includes NVIDIA's Physical AI Data Factory skills and is open to robotics development teams using Nebius AI Cloud, with computing and storage charged according to usage — meaning the barrier to entry for other 3PLs and robotics vendors is a metered bill, not a proprietary contract.

That openness matters. If the simulation-to-deployment loop proves durable in Zalando's returns environment, expect rival fulfillment operators to replicate the pattern across their own high-variability processes, from mixed-SKU picking to unsorted inbound freight.