Overview

Training robots for real-world tasks requires more than code. It requires high-fidelity, synchronized data captured directly on production hardware. The UR AI Trainer, developed by Universal Robots (a Teradyne Robotics company) in collaboration with Scale AI, is built to do exactly that. At its core is a vision system powered by the Orbbec Gemini 305g, delivering the precise, synchronized RGB and depth data that Physical AI models depend on.

UR AI Trainer at Automate 2026

Orbbec Gemini 305g integrated in UR AI Trainer

The Challenge: Bridging the Lab-to-Factory Gap

Physical AI models are only as good as the data they are trained on. Most training data today is collected on research robots in controlled environments, capturing only visual feedback and lacking the force and motion context needed for contact-rich industrial tasks such as electronics assembly. Transferring models trained in the lab to production environments has remained one of the hardest unsolved problems in robotics.

What is needed is a system that captures high-quality, synchronized motion, force, and vision data directly on the same robots that will ultimately be deployed, closing the gap between research and factory floor.

The Solution: UR AI Trainer

The UR AI Trainer is Universal Robots’ imitation learning platform, developed with Scale AI and deploying on the UR AI Accelerator platform. The system uses a leader-follower setup: a human operator physically guides a smaller UR3e robot through a task, while a larger UR7e follower robot mirrors the motion in real time. Force torque sensors in all robots provide haptic feedback, so the operator can feel how the follower interacts with the environment, enabling more natural and precise demonstrations of contact-rich tasks.

To capture the vision side of this data pipeline, Teradyne Robotics integrated four Orbbec Gemini 305g stereo 3D cameras in two configurations:

Two scene-mounted cameras: one looking down, one looking out at the workspace, providing broad environmental context.

Two wrist-mounted cameras: positioned directly on the robot’s end-effector, capturing close-range interaction between the gripper and objects at the manipulation point.

The GMSL interface on the Gemini 305g supports up to eight camera streams at 60 Hz directly into Jetson video memory. In this setup, all four cameras stream simultaneously at 60 Hz in full resolution, a pipeline that would not be achievable over USB.

Camera streams are displayed live on UR’s PolyScope X teach pendants, giving operators multiple angles of the task in real time. Synchronized motion, force, and vision data are recorded together by Scale AI software, producing structured datasets ready for Vision-Language-Action (VLA) model training, including NVIDIA’s GR00T open VLA model.

Why Orbbec Gemini 305g

The Gemini 305g is the GMSL2 variant of Orbbec’s Gemini 305 wrist-mount stereo vision camera, selected for this application across four key dimensions: its compact 42 mm x 42 mm footprint and 116 g weight fit directly on a cobot wrist without affecting payload; its GMSL2 interface enables four simultaneous 60 Hz camera streams directly into Jetson video memory over bend-resistant cables suited for arms in motion; its 4 cm minimum operating range and accuracy reaching up to sub-millimeter at 15 cm cover the near-field distances of real manipulation tasks; and its global shutter sensors with hardware and software multi-camera sync deliver clean, frame-aligned RGB and depth data across all four viewpoints simultaneously.

Orbbec Gemini 305g close-up on cobot wrist

Results

The UR AI Trainer represents a fundamental shift in how robots learn. Rather than relying on data collected in isolated research labs, it generates high-fidelity training datasets directly on the same production hardware that will be deployed on the factory floor, closing the lab-to-factory gap that has held back Physical AI at scale.

With Orbbec Gemini 305g providing four synchronized camera streams at 60 Hz across both scene and wrist perspectives, the system captures the full picture of every human demonstration: where objects are, how the gripper approaches them, and how contact unfolds in real time. Scale AI software structures and manages the recorded datasets, combining synchronized robot motion, force torque data, and multi-viewpoint RGB and depth video into the multimodal datasets that VLA models need to learn contact-rich tasks reliably.

The result is a data pipeline that turns human demonstrations into deployable robot intelligence, on real hardware, at production scale.

About Teradyne Robotics

Teradyne Robotics is a global leader in advanced robotics solutions, dedicated to revolutionizing manufacturing processes through innovation in collaborative and mobile robotic technology.

Teradyne Robotics companies, Universal Robots and Mobile Industrial Robots (MiR) empower businesses of all sizes to enhance operational efficiency by integrating the power of machines with human talent. Our comprehensive range of solutions enables companies to optimize manufacturing processes, leading to improved product quality, and increased productivity, while greatly improving worker safety.

Teradyne Robotics is a division of Teradyne, Inc. (NASDAQ: TER), a leading provider of automated test equipment and advanced robotics. For more information, visit teradyne.com. Teradyne® is a registered trademark of Teradyne, Inc. in the U.S. and other countries.

Orbbec Product

Gemini 305g — Lightweight Close-Range Stereo Vision Camera for Robotic Wrist Mounting, GMSL2 Interface

Discover Orbbec’s Robot-Free Data Collection Platform