Teradyne Robotics and Sereact to bring AI to retail and logistics with Cortex and cobots

Teradyne Robotics and Sereact announced a partnership on October 21, 2026, to advance industrial AI solutions for retail, logistics and manufacturing. The companies will run Sereact’s physical AI platform, Cortex, on Universal Robots arms for tasks involving variable goods and packaging, including returns processing and piece picking.

For tasks such as returns processing and piece picking, where goods and packaging vary, the companies outlined an approach combining Cortex with Universal Robots arms. Sereact says the same model can be used for multiple tasks without retraining for each item or task. The scope of its applicability and its performance in real-world settings will help determine the potential for automation in retail and logistics.

The partnership centers not on the automation capabilities of the robot arm alone, but on the concept of using Cortex as a common model that can handle different goods and packaging. The targeted applications include returns processing and piece picking. Sereact says the model does not require retraining for each item or task. Publicly available information does not show how broadly this will hold up in real-world operations; handling exceptions and the amount of human intervention will be key points of evaluation. The technical framework consists of training in simulation, additional training through reinforcement learning, benchmarking on Teradyne Robotics hardware, in-cell inference, and development and updates at the site level. It is designed to connect model development and validation with on-site execution, rather than simply placing an AI model on the robot. The use of NVIDIA development platforms and world models also connects to the work on infrastructure for data generation, training, evaluation and deployment covered in NVIDIA adds training, evaluation and deployment tools for world models to Cosmos (published October 22, 2026). However, the platform expansion discussed in that article should not be treated as equivalent to specific deployment results for Cortex. The entry point for robot development also differs. AWS releases a toolchain for robot development (published October 23, 2026) covered the release of a developer toolchain. This announcement, by contrast, is a partnership to combine Universal Robots arms with Cortex for applications such as returns processing, representing a different stage: implementation in workplace tasks rather than provision of a development environment. If a common model expands across multiple robots and tasks, how broadly the benefits of model updates can be shared will be one measure of how it differs from device-by-device configuration. Unresolved questions include throughput at each site, uptime, recognition and grasping failure rates, how performance is verified after model updates, and the conditions for safety validation. The partnership announcement outlines training, evaluation and inference, but does not specify these operational metrics or the limits on applicable goods and tasks. It will be important to determine how much work the demonstrated setup can cover in live operations and how it handles exceptions.

The companies said they will demonstrate returns processing at NVIDIA GTC Berlin 2026, running Cortex’s Vision-Language-Action (VLA) model on two Universal Robots UR12e-1300 arms. Cortex is trained using NVIDIA Isaac Sim and NVIDIA Cosmos, then further trained with reinforcement learning using NVIDIA Isaac Lab and benchmarked on Teradyne Robotics hardware before deployment. In operation, NVIDIA Jetson Thor will be used for in-cell inference, while NVIDIA DGX Spark will support site-level development, validation and model updates. Sereact’s CEO and co-founder said the same production model can be used for returns processing, replenishment and kitting, without retraining for each item or task. The company also said Cortex can handle mixed products, unfamiliar items and a range of packaging.