Sereact, NVIDIA, Nebius and Zalando close the training loop for the Cortex robotic brain

The loop is closed: Sereact, NVIDIA, Nebius and Zalando industrialize robot training in production

🔎 Robotic MLOps finally leaves the lab

On October 21, 2026, a press release goes almost unnoticed between two language model launches. It shouldn't. Sereact, NVIDIA, Nebius, CEVA Logistics and Zalando announce the production deployment of a closed training loop for Cortex, Sereact's vision-language-action (VLA) model, deployed on Zalando returns processing stations hosted in CEVA Logistics fulfillment centers (GlobeNewswire).

Why does it matter? Because so-called "embodied AI" robotics has until now suffered from a structural problem that any ML engineer will immediately recognize: no data loop. A robot makes a mistake, nobody collects the failure in a structured way, nobody feeds it back into training. The model stays frozen, the same errors repeat for weeks, and every fix requires a manual engineering cycle.

What changes with this announcement is that this cycle is now automated end to end: real failures from the stations become data, that data feeds a massive simulation, the simulation feeds post-training, and the updated model goes back to field testing. An error observed on a Monday can end up as a candidate policy tested within days instead of weeks.

And the most interesting detail of the whole announcement is neither the robot nor the GPU: it's a coding agent that drives the loop.

The Essentials

• Closed loop in production: video and cinematic episodes from Sereact return stations are captured, stored on Nebius, transformed into simulated variants via NVIDIA Isaac Lab and NVIDIA Cosmos, then fed back into Cortex post-training on Blackwell GPUs.
• An agent drives the pipeline: a coding agent reads the NVIDIA Cosmos Data Factory blueprints and Nebius Workbench skills — versioned instruction files — to orchestrate each iteration, with a GPU cost preview before each run and human approval beyond a threshold.
• First real-world deployment: Zalando returns handled by dual-arm robotic stations at CEVA Logistics, in Germany and Poland.
• Open reference architecture: the full pipeline is available today on Nebius AI Cloud, EU regions, with usage-based billing (compute + storage), with no separate fees for the Workbench.
• The pattern that matters: this isn't a model, it's a continuous improvement infrastructure for physical AI — the equivalent of what MLOps did for NLP between 2019 and 2022.

Recommended tools

If you want to reproduce this type of pipeline for physical AI or robotic simulation, here are the building blocks mentioned in the announcement and how they fit in:

For teams that want simple hosting of their experimentation stack before scaling up to dedicated GPUs, a host like Hostinger does the job for the web and API parts, but obviously doesn't replace a GPU cloud for training.

How the closed loop works — Direct answer

In one sentence: every robot failure becomes training data, with no manual intervention between capture and post-training. Here is the pipeline as described by Sereact (source).

Step 1: capture of real episodes

Sereact's dual-arm return stations, installed at CEVA Logistics on behalf of Zalando, record their work episodes: video streams from cameras and kinematic data from the arms. Every attempt to grasp a returned garment, every grasp error, every unusual object becomes a timestamped episode.

This is the point most robotics projects have missed until now. Field data exists — it's simply thrown in the trash for lack of infrastructure to recycle it.

Step 2: storage and ingestion

The episodes are sent to Nebius's object storage. Nothing exotic here, and that's precisely the point: the pipeline relies on mundane components (object storage, EU regions, usage-based billing), not on locked-in proprietary infrastructure.

Step 3: simulation and augmentation

This is where NVIDIA comes in. Isaac Lab generates thousands of environment variants from the real episodes: object positions, station layouts, grasp configurations. Then NVIDIA Cosmos, the world foundation model, re-renders these variants with a level of physical realism that matters enormously in returns processing: the state of the garments (crumpled, folded, tangled), the packaging, the lighting conditions.

In the details of the previously published Sereact/Nebius partnership, the pipeline describes seven steps: ingestion, simulation/augmentation (Isaac Lab generates up to 10,000 environment variants), dataset versioning with an 80/20 split, training on 8,000 environments, evaluation on 2,000 hold-out environments with a rubric scored by a VLM, a threshold gate, then test deployment (Sereact). The October 2026 announcement is this pipeline going from paper to production.

Step 4: post-training on Blackwell

Cortex's post-training runs on NVIDIA Blackwell GPUs, in Nebius's EU regions. Simulation runs on RTX-class GPUs, training on Blackwell and B300/GB300 configurations — a separation that makes economic sense, since rendering variants is massively parallelizable on cheaper hardware.

Step 5: evaluation on hold-out variants

An often-overlooked yet decisive point: the updated model is scored on hold-out variants — environments never seen during training — against a threshold defined by Sereact. If the score passes, test deployment. If not, targeted data augmentation and retraining. This is a real quality gate, not a deployment based on a gut feeling.

The real story: an agent that drives the training infrastructure

The most under-commented part of the announcement is the role of the coding agent. It's not a human who launches the jobs, adjusts the parameters, and monitors costs. It's a software agent that:

1. Reads the NVIDIA Cosmos Data Factory blueprints — versioned instruction files that function as agentic skills: the agent doesn't "know" the pipeline by heart, it reads it from a versioned repository, the way a developer reads a runbook.
2. Reads the Nebius Workbench skills for infrastructure operations.
3. Presents a GPU cost forecast before each run — and that's where human governance comes in: above a defined cost threshold, human approval is required.

This pattern — versioned instructions as the interface between an agent and an infrastructure — is exactly the one emerging everywhere in agentic in 2026. We saw it with Google Research's HARNESS, which lets agents rewrite the system around them: the question is no longer "which model", but "which versioned interface does the model read in order to act".

Why is this clever here? Because a robotic training pipeline is expensive and slow. A human orchestrating by hand introduces bottlenecks and handling errors. An agent with cost forecasting and governance gates turns the pipeline into an internal service available around the clock. Nebius also reports that Physical AI Workbench customers see up to 40% less engineering time (Nebius).

My take, no sugarcoating: the cost threshold with human approval is the most mature detail of the entire architecture. Many agentic projects fail precisely because they lack a budget guardrail. Here, it's built into the pipeline from the start.

Why returns processing is the perfect use case

Zalando's choice is no accident, and it deserves a closer look. Returns processing is probably the worst grasping nightmare that exists in logistics:

• Garments arrive in unpredictable physical states: crumpled, folded into a ball, still inside a bag, tangled up with other items.
• The variability is nearly infinite: every season brings new materials, cuts, and packaging.
• The volume is enormous — in online fashion, return rates can exceed 30 to 50% depending on the category, a figure documented for years in European e-commerce studies.

A VLA model trained once on static data ages poorly in the face of this variability. Hence the absolute necessity of a closed loop: it's not a nice-to-have, it's a condition for the product's survival. Sereact, founded in 2021 in Stuttgart and having raised more than $145M in total, has made this use case its Trojan horse — and the deployment at CEVA for Zalando gives it what few robotics startups have: a real, continuous, and industrially significant data stream.

It's also a signal for the rest of the logistics sector. If the loop works on returns — the hardest case — it transfers to simpler cases (standard pick-and-place, palletizing) with very little adaptation.

Open architecture as a strategy: Nebius publishes the reference architecture

Another thing not to miss: the pipeline is published as an open reference architecture on Nebius AI Cloud, EU regions, with usage-based billing on compute and storage, with no separate fees for the Workbench (CEVA Logistics).

It's a triple-play strategic move:

1. For Nebius: every European robotics team that reproduces the architecture becomes a Nebius GPU consumer. The open source Workbench is the horse, the cloud is the hay.
2. For NVIDIA: every reproduction consolidates the Isaac + Cosmos ecosystem as the de facto standard for physical AI. We're far from the years when every lab reinvented its own simulation stack.
3. For the ecosystem: a documented reference architecture, tested in production at a leading logistics company, is worth a thousand research papers.

This move fits into the open source dynamic that NVIDIA has been pushing since GTC Berlin, where Isaac GR00T became the ready-to-use reference humanoid and where the company joined forces with Hugging Face in LeRobot to make GR00T 1.7, Isaac Teleop, and Cosmos 3 accessible. The logic is consistent: NVIDIA isn't just selling chips, it's selling the complete software infrastructure that makes those chips indispensable — and the more the ecosystem opens up, the stronger the hardware lock-in becomes.

What this changes for European robotics teams

If you run a robotics team or an applied lab in Europe, three concrete implications:

Field data becomes a strategic asset. Companies operating robots in production now hold an advantage that is nearly impossible to replicate: a continuous stream of real episodes. If you haven't planned for capturing and versioning your episodes, you're throwing away your most valuable asset.

Sim-to-real becomes an engineering discipline, not a research project. With Isaac Lab for variant generation and Cosmos for re-rendering, the boundary between "real data" and "simulated data" becomes a pipeline parameter, not an academic debate. Teams that master this mix will produce more robust models than those stubbornly collecting everything in the real world.

Agentic orchestration changes team profiles. The role of the robotics ML engineer is shifting toward designing skills, evaluation gates, and governance thresholds. It's the same skills shift observed in software MLOps: fewer manual scripts, more design of self-improving systems.

To place this development in the broader landscape: Sereact is not alone in the data → world model → policy loop. GreenWAM, Sber's world model, drives 20 robot families with a single model follows a similar philosophy, and open foundations like Cosmos 3 and Isaac GR00T mark what several observers are calling the "ChatGPT moment" of robotics. The difference: Sereact is one of the first to run the loop in production, at a customer site, on a real business use case — not in a demo.

What the announcement doesn't say

A bit of journalistic distance is in order, as the press release leaves some gray areas:

• Performance metrics are not published. How many episodes are collected per week? What is the rate of improvement in hold-out score per iteration? What is the actual delay between error detection and test deployment? Nothing quantified in the press release beyond the "up to 40% less engineering time" reported by Nebius, which remains a measure of effort, not of quality.
• The quality gate depends on a VLM. Having a vision-language model score the success of a manipulation is elegant, but it introduces a judge that is itself imperfect. Sereact does not detail the calibration of this rubric, nor its rate of disagreement with human evaluation.
• Generalization beyond feedback mode remains to be proven. The pipeline is validated on a single domain. Clothes have physical properties (deformability) that make simulation particularly difficult — it's a case where Cosmos brings a lot, but also a case that doesn't prove everything.

None of this invalidates the announcement. But let's distinguish what is demonstrated — a loop running in production on a real industrial case — from what is suggested — a general solution for physical AI.

❌ Common Mistakes

Mistake 1: Believing the closed loop replaces real-world data collection

Sereact's pipeline doesn't generate data out of nothing: it amplifies real episodes via simulation. Without an initial stream of field episodes, Isaac Lab has nothing to vary and Cosmos nothing to re-render. The solution: design episode capture as a first-class product feature, with versioning and a train/hold-out split from day one — the 80/20 split described by Sereact is a good starting point.

Mistake 2: Deploying agentic orchestration without a cost guardrail

Letting an agent launch GPU runs without budget forecasting or an approval threshold is the fastest way to receive a surprise five-figure bill. The solution is in the announcement itself: cost forecast before each run, mandatory human approval above a threshold. Copy this pattern as-is.

Mistake 3: Evaluating on training data (or on variants that are too similar)

The classic sim-to-real trap is overfitting to the simulated variants: the model excels in simulation and collapses in the real world. The solution: strictly separated hold-out environments, generated after training, and a deployment threshold defined by business operators — not by the ML team.

❓ Frequently Asked Questions

What exactly is Cortex?

It's Sereact's vision-language-action model: a model that takes visual observations and instructions as input, and produces robotic actions. It powers the dual-arm returns processing stations deployed at CEVA Logistics for Zalando.

Is the closed loop available to other companies?

Yes, that's the whole point of the reference architecture published on Nebius AI Cloud in EU regions. It's billed on a usage basis (compute + storage) with no separate fees for the Workbench, which is open source. But reproducing the pipeline assumes you have robots in production and episodes to ingest.

What exactly is NVIDIA's role in this partnership?

NVIDIA provides the simulation and world modeling layer: Isaac Sim, Isaac Lab, Cosmos, and the Physical AI Data Factory skills. Nebius provides the execution infrastructure in EU regions with RTX-class GPUs for simulation and Blackwell for training.

Why is EU regulation an argument here?

Data captured in warehouses (worker images, logistics flows) stays within EU regions, which simplifies GDPR compliance. For Zalando and CEVA, this was most likely a non-negotiable condition of the deployment.

How much does a full iteration of the loop cost?

Not disclosed. The pipeline includes a GPU cost forecast displayed before each run, which suggests the cost varies significantly depending on the volume of generated variants and the duration of post-training. This is precisely why the human approval threshold exists.

Does this make warehouse operators obsolete?

No, and that's not the point. The stations handle one specific task (returns sorting) in a workflow where humans remain present for unhandled cases. Returns automation primarily addresses a problem of volume and turnaround time, not wholesale replacement.

✅ Conclusion

Sereact, NVIDIA, Nebius, CEVA and Zalando's closed loop doesn't revolutionize a model — it revolutionizes the model's lifecycle, and that's far more sustainable. Where applied robotics was still tinkering with manual fixes, an agent now pilots, with cost governance, a pipeline that turns every field failure into a measured improvement of the robot's brain. If you work on physical AI, the lesson to remember comes down to one sentence: stop thinking "model," start thinking "loop" — and go read the reference architecture before your competitors do.