How they use Nebius AI Cloud

Sereact runs a live robot learning pipeline on the Nebius Physical AI Workbench, initially for robots replenishing store shelves. The workflow connects real robot episodes with simulation, augmentation, policy training and evaluation before a real-world robot test. AI agents assemble and operate the pipeline using NVIDIA Isaac, Cosmos and Cosmos Data Factory.

Video and kinematics land in Nebius object storage. The pipeline generates 10,000 environment variants, trains policies on 8,000 environments and evaluates them on 2,000 held-out environments. Failed evaluations trigger targeted augmentation, retraining and another evaluation. Nebius supplies RTX-class GPUs for simulation and augmentation and Blackwell GPUs for training in EU data centers. The announcement describes broader scenario coverage and a shorter learning loop as the partnership's aims.

Sereact also announced a Cortex training loop for Zalando returns handling at selected CEVA Logistics fulfillment centers. Returns-station episodes enter Nebius object storage. Isaac Lab and Cosmos generate variants across garment states, packaging, materials and lighting to post-train Cortex on Blackwell GPUs. An AI agent operates the loop and previews GPU cost before each run. Sereact engineers set goals and approve runs above a set cost limit. Sereact says tested policy updates can return within days, compared with weeks previously.