Teradyne Robotics and Sereact pair natural-language processing with industrial physical AI by running Sereact's Cortex platform on Universal Robots cobot arms. Cortex is a Vision-Language-Action model that takes plain-language instructions plus camera input and turns them into robot actions.
Universal Robots says the companies announced the partnership in Berlin on Oct. 21, 2026 to scale physical-AI solutions across retail, logistics and manufacturing partnership announcement. The target is unstructured work such as returns handling and piece-picking.
Table of Contents
- What does the language command do?
- Which warehouse tasks fit this approach?
- Where is it already deployed?
- How is it trained and where are the limits?
What does the language command do?
Cortex processes natural-language instructions alongside visual input and produces corresponding actions through discrete plans and tokenized actions. Sereact describes this design in its Cortex overview technical overview. Instead of task-by-task code, an operator gives a plain-language goal such as "Clear the clutter from this bin and repack items by size." The system grounds that goal into perception and control.
The operator states intent, while the model selects grasps, motions and sort actions. In practice, this shifts setup from programming to instructing. Staff describe the desired outcome for the bin, tote or return. The robot interprets objects in view and executes the steps.
Which warehouse tasks fit this approach?
The joint solution focuses on mixed goods, unknown items and varied packaging. It handles these without product-specific teaching or a fixed item catalog. That enables zero-shot grasping and sorting for returns and piece-picking. A cell can pick an item it never saw before.
It can also separate sizes, product types or packaging styles in one flow. Otto Group reporting on the approach links it to returns processing without per-item training returns deployment report. That matters when inbound volume changes daily. Fashion returns are a clear example because color, fabric, size and bagging vary widely.
Where is it already deployed?
Retailers and logistics providers are the first users. The use case is autonomous fashion-returns processing at fulfillment scale. TechTimes reports that CEVA Logistics activated dual-arm Cortex and Universal Robots systems on Oct.
7, 2026 at Zalando centers in Greven, Germany and Świebodzin, Poland deployment report. Each cell processes hanging and bagged garments without item-specific setup. For site managers, the test is throughput during peak returns, not demo picking.
How is it trained and where are the limits?
Cortex is trained and improved with NVIDIA Isaac Sim, Isaac Lab reinforcement learning, Cosmos world models and Jetson Thor edge inference. Live picks feed back into simulation. That loop builds skill before deployment and refines behavior after. Simulation covers many grasps, slips and package shapes.
Production data corrects what simulation misses. Fulfillment centers remain difficult because reflective and deformable objects challenge vision and grasping. Sereact says reliable use still needs real operational data plus depth and mask perception and high-level planning layers. Proven returns cells are therefore offered as Robotics-as-a-Service for scaling from Europe to North America:.
- start with one returns cell with defined inbound mix
- connect live picks to retraining and monitoring
- add depth sensing and planning checks for deformables
- expand to more cells only after stable uptime
