CEVA Logistics and Zalando are working with Sereact, NVIDIA and Nebius on a system that uses warehouse data and simulation to improve robots handling e-commerce returns. The companies say the approach could cut training cycles from weeks to days, a useful promise for warehouses where a returned jumper can arrive in almost any state except neatly folded.
Watch Desk analysis
What happened
The collaboration, announced on 22 October, targets Zalando returns operations at selected CEVA fulfilment centres. Sereact’s Cortex robotic AI system collects video and movement data during warehouse work. Nebius provides cloud infrastructure, while NVIDIA’s Isaac tools and Cosmos generate simulated variations of handling scenarios, including different garment conditions, packaging and lighting.
The expanded datasets are used to train Cortex on NVIDIA Blackwell GPUs. Updated versions are tested against performance criteria set by Sereact: successful versions move to real-world trials, while failures trigger further simulation and training. Engineers set training goals and approve operations that exceed predefined computing-cost limits. The companies also say the training pipeline is available as a reference architecture through Nebius’ Physical AI Workbench.
Why it matters
Returns are awkward work for robots: products may be crumpled, unpackaged or missing labels, and the stock itself changes. Training with real warehouse data and simulated variations could help a robot encounter more of that mess before it is asked to handle the real thing.
The claimed shift from weeks to days would make it easier to adapt systems as products and operations change. That is the companies’ stated aim, not a measured result supplied here. The deployment offers a concrete test of whether simulation can speed up improvements without leaving the actual warehouse behind.
Our read
The interesting part is the feedback loop, not simply that several familiar AI names are in the room. Real handling data informs simulation, simulation informs training, and operational testing sends failures back for another go. That is a more practical recipe for warehouse robotics than hoping a model has met enough imaginary parcels.
For operators, the appeal will depend on whether faster training translates into reliable handling across changing returns, and what it costs in cloud compute and human oversight. The initial deployment is a use case to watch, not yet proof that the promised timetable has arrived.
What to watch
- Whether the system reaches real-world trials and how it performs on varied returns.
- Whether training cycles actually fall from weeks to days.
- How much computing the loop uses, and how often engineers need to intervene.
Discussion spark: Should warehouse robots be trained mainly on real operational data, or is simulation the better route to handling the mess and variety of returns?
Sources and evidence
- Source update (26 October 2026, 01:30 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.