Sereact, NVIDIA, Nebius, CEVA Logistics and Zalando develop AI training loop for robotic returns handling
Sereact, NVIDIA, Nebius, CEVA Logistics and Zalando have joined forces to develop a closed-loop AI training system designed to improve robotic returns handling in e-commerce logistics. The initial deployment will focus on Zalando’s returns operations at selected CEVA Logistics fulfilment centres.
The system combines NVIDIA’s simulation and AI technologies, Nebius’ Physical AI Workbench and Sereact’s Cortex robotics platform. The aim is to shorten the time required to retrain and validate robots, allowing operational errors to be addressed through tested policy updates within days rather than weeks.
Returns processing presents a significant automation challenge because products arrive in varying conditions, with differences in folding, packaging, labels and materials. Robots must pick up items, identify them, assess their condition and sort them, despite constantly changing product assortments.
Linking real-world data with AI simulation
Under the new workflow, video and movement data recorded at returns stations are stored in Nebius’ object storage. NVIDIA Isaac Lab uses the data to generate simulated variations, while NVIDIA Cosmos recreates different garment conditions, packaging, materials and lighting scenarios.
The resulting datasets are used to retrain Sereact’s Cortex system on NVIDIA Blackwell GPUs. Updated policies are evaluated against predefined performance thresholds before being tested in real-world warehouse operations. If a policy fails, the system generates additional simulated scenarios targeting the cases it did not handle successfully, and the training cycle begins again.
An AI agent coordinates the process, from launching simulation and training stages to evaluating results and identifying cases that require further testing. Sereact engineers define the objectives and approve runs that exceed specified computing cost limits.
Roles across the partnership
Sereact contributes its Cortex vision-language-action model, dual-arm robotic returns stations and operational data. NVIDIA provides Isaac Sim, Isaac Lab, Cosmos and Physical AI Data Factory skills, while Nebius supplies the Physical AI Workbench framework, shared data infrastructure and computing resources through its European data centres.
CEVA Logistics hosts the robotic stations and manages e-commerce logistics operations for Zalando, which provides the returns-handling use case for testing the updated AI policies.
The closed-loop pipeline is available as a reference architecture through Nebius AI Cloud, with NVIDIA’s Physical AI Data Factory skills pre-installed. The system operates in Nebius’ European regions, with customers paying for the computing and storage resources they use.