Sereact and Nebius Add a Closed Training Loop for Warehouse Robots

Sereact, NVIDIA and Nebius announced a closed training loop for warehouse robots on October 21, using Zalando returns handled by CEVA Logistics as the first deployment. The workflow connects recorded robot episodes to simulation, model updates and testing. Its useful change is how quickly a difficult case can become a tested improvement.

What changed for Sereact Cortex

Cortex is Sereact's vision-language-action model, the software that connects what a robot sees and understands to its movements. The joint announcement describes a workflow intended to shorten policy-update cycles from weeks to days. That is a company claim about iteration time, not a published improvement in picking accuracy.

The warehouse operation already existed. Sereact and its partners announced live returns processing on October 7, with dual-arm systems in Greven, Germany, and Świebodzin, Poland. Those systems grasp, identify and sort returned fashion items under a Robotics-as-a-Service arrangement.

The October 21 development adds a way to improve those systems using the situations they encounter. It should not be read as a new robot launch or the first deployment of Cortex.

How the closed training loop works

A recorded episode combines video with information about the robot's movement. Episodes enter Nebius object storage, Isaac Lab generates simulated variations, and NVIDIA Cosmos changes their visual conditions. The resulting versioned dataset is used to post-train Cortex on Blackwell GPUs.

The next step is evaluation on held-out variations. Sereact sets the passing threshold. A passing policy proceeds to a physical test; a failed policy sends the workflow back toward additional examples of the cases it missed.

Passing a simulation gate earns a real-world test, not an automatic claim of production readiness. That boundary matters because a model can become better at the scenarios used for training without becoming reliable on the next unfamiliar return.

Consider a garment arriving with packaging partly covering it. Generating many versions of that situation could make the failure easier to study. But useful variation needs to preserve the physical properties that affect the grasp. More images alone do not establish that the robot has learned the right movement.

What Nebius Physical AI Workbench provides

The Nebius announcement describes new integration with NVIDIA Cosmos Data Factory, including agent-readable instructions and ready-made workflows for preparing training data. Its Sereact collaboration applies that infrastructure to returns handling.

Workbench provides the surrounding workflow system rather than replacing Cortex. The public repository documents a command-line interface, supported Python interfaces, containerized jobs and shared object storage. A coding agent can plan and operate workloads through those tools, while engineers can use the same interfaces directly.

The repository also separates different kinds of evidence. Starting a container, producing a checkpoint and demonstrating successful robot behavior are different outcomes. It explicitly cautions that a smoke test does not prove a policy has converged.

For developers evaluating the stack, that is a useful distinction. Before adopting a reference workflow, check what its recorded test actually established and whether its input data, hardware and model access match the proposed workload.

Availability, pricing and approval limits

The partners say the closed-loop reference architecture is available on Nebius AI Cloud in EU regions, with charges for compute and storage. The announcement does not give a fixed price per training cycle. Engineers approve runs that exceed a configured spending limit.

Workbench's code is Apache-2.0, but third-party models, software and datasets retain their own terms. Its setup documentation also distinguishes configuration checks from GPU reservation and model-access checks. A working local command is only one part of getting a workload ready.

That makes cost control part of the design. If an agent keeps generating examples until a policy passes, a team needs to decide which failures are worth another training run and how much evidence is sufficient before testing hardware.

The useful measure is the next deployment

For a robotics team, the attraction is a shorter path from an observed failure to an evaluated change. The commercial test comes later: fewer repeated exceptions at the station, with the cost of generating and testing those improvements included.

A faster training loop creates more opportunities to improve a robot. The important evidence will be whether those opportunities produce better behavior on returns the system has not seen before.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn complex AI launches and engineering updates into clear stories that developers and founders can understand and share.

Book a 30-minute call.