CEVA closes training loop for Sereact’s robot brain
0Sereact, NVIDIA, Nebius, CEVA Logistics and Zalando today announced a closed training loop for Cortex, Sereact’s robot brain.
The first deployment is Zalando returns handling in selected CEVA Logistics fulfillment centers. NVIDIA provides Isaac Sim, Isaac Lab, Cosmos and its Physical AI Data Factory skills. Operated through the company’s EU data centers, Nebius runs them on the Physical AI Workbench, its open-source framework for physical AI. Sereact brings Cortex and the episodes its robots record on the returns line. What a robot gets wrong on Monday can come back as a tested policy update within days, where it used to take weeks.
Returns are among the hardest warehouse processes to automate. Each item arrives in whatever state the customer left it: refolded or crumpled, in its polybag or loose, with or without tags. A robot has to pick it, identify it, check its condition and sort it, while the assortment changes every season. Cortex learns from real picks, but real picks only cover items that have already passed through a station. Improving the policy meant collecting episodes on site, retraining, testing on the robot and starting over. Each cycle took weeks and covered only the cases a team could stage by hand.
The new loop runs as one workflow on one shared data layer. Episodes from the returns stations, video plus kinematics, land in Nebius object storage. NVIDIA Isaac Lab turns them into thousands of simulated variants, and NVIDIA Cosmos re-renders those across garment states, packaging, materials and lighting. The versioned dataset post-trains Cortex on NVIDIA Blackwell GPUs. The new policy is scored on held-out variants against a threshold Sereact sets. A policy that passes goes to the stations for a real-world test. One that fails triggers targeted variants for the cases it missed, and the loop trains again.
An AI agent builds and runs the loop. NVIDIA’s Physical AI Data Factory skills are open, versioned instruction files that a coding agent reads directly, and Nebius adds Workbench skills for simulation, training, evaluation and the gate. The agent previews GPU cost before each run, launches the stages, checks the gate and, after a failed run, decides which cases to augment. Sereact engineers state the goal, such as “validate the policy for the autumn assortment”, and approve any run above a set cost limit.
Sereact brings Cortex, its vision-language-action model and robot brain, along with the dual-arm returns stations and the real episodes they record.
NVIDIA brings Isaac Sim and Isaac Lab for simulation, Cosmos world foundation models for re-rendering, and the Physical AI Data Factory skills that let an agent operate the loop.
Nebius brings its open-source framework, Physical AI Workbench, with shared object storage, CLI, SDK and format contracts at every handoff, and the GPUs behind it: RTX-class for simulation and Blackwell for training, all operated through EU data centers.
CEVA Logistics manages eCommerce logistics operations, including returns handling, for Zalando and hosts the Sereact stations in its fulfillment centers.
Zalando provides the use case. Its returns are where every new policy has to prove itself.
The closed-loop pipeline is available today as a Workbench reference architecture, with the NVIDIA Physical AI Data Factory skills pre-installed, for any robotics team on Nebius AI cloud. It runs in Nebius EU regions, and customers pay only for the compute and storage they use.
Sereact, NVIDIA, Nebius, CEVA Logistics and Zalando today announced a closed training loop for Cortex, Sereact’s robot brain.
The first deployment is Zalando returns handling in selected CEVA Logistics fulfillment centers. NVIDIA provides Isaac Sim, Isaac Lab, Cosmos and its Physical AI Data Factory skills. Operated through the company’s EU data centers, Nebius runs them on the Physical AI Workbench, its open-source framework for physical AI. Sereact brings Cortex and the episodes its robots record on the returns line. What a robot gets wrong on Monday can come back as a tested policy update within days, where it used to take weeks.
Returns are among the hardest warehouse processes to automate. Each item arrives in whatever state the customer left it: refolded or crumpled, in its polybag or loose, with or without tags. A robot has to pick it, identify it, check its condition and sort it, while the assortment changes every season. Cortex learns from real picks, but real picks only cover items that have already passed through a station. Improving the policy meant collecting episodes on site, retraining, testing on the robot and starting over. Each cycle took weeks and covered only the cases a team could stage by hand.
The new loop runs as one workflow on one shared data layer. Episodes from the returns stations, video plus kinematics, land in Nebius object storage. NVIDIA Isaac Lab turns them into thousands of simulated variants, and NVIDIA Cosmos re-renders those across garment states, packaging, materials and lighting. The versioned dataset post-trains Cortex on NVIDIA Blackwell GPUs. The new policy is scored on held-out variants against a threshold Sereact sets. A policy that passes goes to the stations for a real-world test. One that fails triggers targeted variants for the cases it missed, and the loop trains again.
An AI agent builds and runs the loop. NVIDIA’s Physical AI Data Factory skills are open, versioned instruction files that a coding agent reads directly, and Nebius adds Workbench skills for simulation, training, evaluation and the gate. The agent previews GPU cost before each run, launches the stages, checks the gate and, after a failed run, decides which cases to augment. Sereact engineers state the goal, such as “validate the policy for the autumn assortment”, and approve any run above a set cost limit.
Sereact brings Cortex, its vision-language-action model and robot brain, along with the dual-arm returns stations and the real episodes they record.
NVIDIA brings Isaac Sim and Isaac Lab for simulation, Cosmos world foundation models for re-rendering, and the Physical AI Data Factory skills that let an agent operate the loop.
Nebius brings its open-source framework, Physical AI Workbench, with shared object storage, CLI, SDK and format contracts at every handoff, and the GPUs behind it: RTX-class for simulation and Blackwell for training, all operated through EU data centers.
CEVA Logistics manages eCommerce logistics operations, including returns handling, for Zalando and hosts the Sereact stations in its fulfillment centers.
Zalando provides the use case. Its returns are where every new policy has to prove itself.
The closed-loop pipeline is available today as a Workbench reference architecture, with the NVIDIA Physical AI Data Factory skills pre-installed, for any robotics team on Nebius AI cloud. It runs in Nebius EU regions, and customers pay only for the compute and storage they use.