New Robotic Returns Training Loop Cuts Update Time
The platform accelerates policy updates for warehouse robots from weeks to days using integrated simulation.
Updated on Oct. 21, 2026 in Robotics
Live Poll
Do you believe the expanded use of AI-driven robotics in warehouses will benefit shoppers?
Sereact, NVIDIA, Nebius, CEVA Logistics, and Zalando have launched a closed training loop for robotic returns handling in Europe. This system integrates robot vision, simulation, and policy training to automate complex tasks that previously required manual intervention.
Why it matters
Returns processing is highly variable due to diverse item conditions and packaging, making automation difficult for traditional systems. By automating these tasks, this architecture reduces the time required for robot policy updates, enabling faster adaptation to changing warehouse conditions.
The system reduces the time required for robot policy updates from weeks to days. It leverages the Physical AI Workbench, running on NVIDIA Blackwell GPUs in Nebius data centers, to iterate on robotic behaviors.
The players
Sereact
A Stuttgart-based robotics company that develops Cortex, an AI-powered robot brain for automated manipulation tasks.
NVIDIA
A technology company providing hardware and software, including Blackwell GPUs and the Isaac simulation platform, for AI and robotics development.
Nebius
An Amsterdam-based cloud platform provider offering AI-optimized infrastructure and storage services.
CEVA Logistics
A global logistics and supply chain manager that operates fulfillment centers for retail partners.
Zalando
A major European online fashion platform serving 62 million active customers across 7,000 brands.
The details
The loop functions by recording real-world episodes from robotic returns stations and storing them in Nebius object storage. These recordings are processed using NVIDIA Isaac Sim and Isaac Lab—tools for virtual environment creation—to generate simulated variants. These variants are then used to retrain and test Sereact’s Cortex robot brain, which powers the dual-arm robotic hardware used in the fulfillment process.
Timeline
October 21, 2026: The closed training loop for robotic returns officially launches.
The Tech Race
This project moves beyond standard automated sorting by integrating real-world warehouse data into a continuous simulation loop. It represents a significant step in the race to deploy 'Physical AI' in environments that require high flexibility, positioning it ahead of rigid, rule-based automation systems.
This system changes how fulfillment centers handle returns, potentially increasing processing speed for Zalando customers. Logistics operators and warehouse managers can now access this reference architecture on the Nebius AI cloud for their own automation projects.
The takeaway
This partnership demonstrates how cloud-native simulation can bridge the gap between static robot programming and the unpredictable nature of warehouse returns. Industry observers should watch for adoption rates of the Physical AI Workbench among other logistics providers as a barometer for real-world reliability.
Further reading
For more on how simulation is transforming industrial automation, explore the Robotics archive.
Live Poll
Do you believe the expanded use of AI-driven robotics in warehouses will benefit shoppers?
