Universal Robots and Scale AI have introduced a system designed to train artificial intelligence models directly on industrial robots, aiming to reduce the gap between research environments and factory deployment. The system, called the UR AI Trainer, enables developers to collect training data on the same hardware used in production settings.
The platform was presented during the GTC 2026 conference in Silicon Valley. It combines Universal Robots’ collaborative robot hardware with Scale AI’s data infrastructure to support the development of AI-driven robotic applications through imitation learning.
According to Universal Robots, the system allows human operators to guide robots through tasks while recording synchronized data used to train machine learning models. In the setup, an operator physically moves a “leader” robot while a “follower” robot replicates the motion in real time. During these demonstrations, the system captures multimodal data including robot motion, force feedback and visual inputs, which can be used to train Vision-Language-Action models.
Anders Beck, vice president of AI robotics products at Universal Robots, said customers require methods for collecting high-fidelity data directly from production-grade robots in order to train models for real-world deployment. He said the system enables synchronized robot and vision data collection on the same platforms used in industrial operations.
The training system incorporates Universal Robots’ direct torque control and force feedback capabilities, which allow developers to capture information about how robots physically interact with objects and environments. The approach is intended to address limitations in robotics AI development, where training data is often gathered on research platforms that differ from the robots used in manufacturing environments.
Scale AI’s software platform manages the capture and structuring of the training data. Ben Levin, general manager of physical AI at Scale AI, said the collaboration enables companies to train, deploy and refine AI models using data collected from operational robots.
The two companies plan to release a large-scale industrial dataset collected from Universal Robots systems later in the year.
Demonstrations accompanying the launch include a smartphone packaging task performed using two pairs of robots. Operators guide two UR3e robots that act as leaders, providing haptic input to control two UR7e follower robots executing the task. Data generated during the demonstration is recorded through Scale AI’s software and can be replayed on the training system.
Another demonstration uses a simulated version of the task built in NVIDIA Omniverse using the Isaac Sim robotics simulation platform. In the simulation environment, users control a virtual bi-manual UR3e system using haptic devices, allowing the task to be trained and tested in a physics-based digital environment.
Universal Robots is also exploring the use of NVIDIA’s Physical AI Data Factory Blueprint to automate the generation of synthetic robotics training data.
In a separate demonstration, Generalist AI showed a robotic foundation model operating on Universal Robots hardware. Two UR7e robots autonomously completed a smartphone packaging workflow designed to illustrate coordinated manipulation and contact-rich handling tasks.
Pete Florence, co-founder and chief executive of Generalist AI, said the demonstration illustrates how data collection and model development can translate into real-world robotic capabilities on industrial platforms.
