Sereact connects robot operations to automated retraining

Sereact has introduced a training and evaluation workflow linking operating warehouse robots with simulation and AI model development, using NVIDIA robotics technology and computing infrastructure from Nebius. The system processes recorded handling attempts from Sereact’s Cortex-powered robots, generates additional simulated scenarios and evaluates updated control policies before they return to physical equipment for testing.

The development, announced on 21 October, involves NVIDIA, Nebius, CEVA Logistics and Zalando. It builds on the deployment of Sereact’s dual-arm returns-handling robots at CEVA facilities in Greven, Germany, and Świebodzin, Poland, which began operating earlier in October. The newly introduced element is the connection between those physical systems and a repeatable training process.

Fashion returns create difficult conditions for robotic manipulation because the same product can arrive in different physical states. Garments may be folded, crumpled, loosely packaged or partly concealed, while labels and accessories can change the appearance of individual items. The range of products also varies as new seasonal assortments enter the returns process.

A robot handling those objects must identify an appropriate contact point, move its gripper towards the object and determine whether the intended manipulation has succeeded. Perception and control systems must respond when the item moves unexpectedly or when its physical characteristics differ from those represented in the training data.

Sereact uses its Cortex vision-language-action platform to support those functions. The software connects information about a scene with actions that can be executed through the robot’s control system. The performance of a particular manipulation depends on the model’s interpretation of the observations as well as the mechanical characteristics of the gripper, the object and the surrounding equipment.

Previously, improving a robotic policy required engineers to collect examples from operating stations, organise additional training and test the revised software on physical hardware. The available recordings were limited to circumstances already encountered by the robots or deliberately recreated by the development team.

The new workflow begins by collecting recorded episodes from returns stations. These include video observations and kinematic information describing how the robot moved during an operation. The recordings are transferred into Nebius object storage, where they can be used by simulation and model development tools operating through a shared data environment.

NVIDIA Isaac Lab generates simulated variations based on those episodes, while Cosmos produces alternative visual representations of garment conditions, packaging materials and lighting. The objective is to increase the range of examples available for training without requiring every physical variation to be staged manually.

This process can be useful where a robot repeatedly encounters a rare failure condition. A recorded unsuccessful grasp provides information about one particular arrangement of the object and robot. Simulation can vary aspects of that arrangement, allowing the development process to test whether an updated policy handles related circumstances more consistently.

The value of those simulated examples depends on their relationship to physical behaviour. A garment may deform or move differently from the representation used by the simulation, particularly where fabric properties, friction or contact forces are not reproduced accurately. An apparent improvement in a virtual environment therefore requires subsequent testing with real objects and mechanical equipment.

The partners combine the recorded episodes and generated variations into versioned training datasets. Cortex can then be updated using NVIDIA Blackwell GPUs hosted by Nebius, with RTX-class computing used for simulation stages. The separated computing resources allow workloads with different processing characteristics to be assigned to suitable hardware.

Dataset versioning is important because changes to a trained policy must be traceable to the information used during its development. If a revised model behaves differently from an earlier version, engineers need to determine whether the change reflects the training material, the training process or another element of the software.

After training, the updated policy is evaluated against examples held apart from the immediate training data. Sereact defines an acceptance threshold, and the candidate must satisfy that threshold before proceeding to physical testing at a returns station.

A failed evaluation can trigger additional simulated cases focused on the circumstances in which the model performed inadequately. The subsequent training run can then incorporate those examples, while the evaluation process provides a mechanism for checking whether the revised behaviour meets the defined criteria.

Testing against separate examples helps identify whether a model has learned behaviour that extends beyond the exact situations used during training. It does not establish complete generalisation across every object or handling condition, particularly when the operating environment changes or previously unseen materials enter the process.

The system is coordinated using NVIDIA Physical AI Data Factory skills and Nebius Physical AI Workbench. The skills provide defined instructions for the simulation, training and evaluation processes, allowing an AI coding agent to initiate the relevant stages and inspect the resulting measurements.

Nebius supplies supporting interfaces, shared storage and format contracts between the stages. These elements reduce the need for a developer to move information manually between separate systems. They also help maintain the relationship between source recordings, generated simulations, trained model versions and evaluation results.

The agent estimates the computing cost before starting a run. Sereact engineers establish the objective and must approve work that exceeds an agreed cost threshold, retaining control over potentially expensive training activity. The software can organise repeated operations, but it does not replace engineering responsibility for the behaviour of the robot.

A candidate policy that passes the defined evaluation threshold proceeds to a physical test, rather than automatically becoming the production controller for every installation. This distinction is important because simulated success may not be reproduced when the robot encounters real object deformation, sensor noise, varying grip conditions or mechanical wear.

The five partners expect the integrated process to reduce some development cycles from weeks to days. That estimate reflects the anticipated benefit of connecting data collection, simulation and evaluation into a common workflow. The announcement does not provide an independently verified comparison of retraining duration, manufacturing throughput or robot reliability across a representative sample of production installations.

The closed-loop pipeline is available through Nebius Physical AI Workbench as a reference architecture, with the NVIDIA skills installed for customers using its European computing infrastructure. Customers pay for the compute and storage consumed, while the suitability of the approach for other robots depends on the data available, the simulation model and the acceptance criteria required by the application.

The programme adds a defined method for investigating manipulation failures and developing revised policies from operational evidence. Its continuing performance will depend on whether the simulated scenarios capture the behaviour of physical objects accurately and whether improvements found during evaluation remain reproducible when the robot returns to the warehouse floor.