Introducing the new UR AI Trainer
AI training on industrial robots.
Many AI models fail to reach production not because of the algorithm, but because the training setup does not reflect industrial reality.
Training on lightweight research arms or vision‑only systems limits what models can learn about contact, compliance, and interaction. When those models are transferred to industrial robots, performance breaks down.
Universal Robots’ AI Trainer is designed to solve this problem at the source.
Force‑aware learning
Integrated force and torque sensing, combined with Direct Torque Control, allows models to learn what the task feels like, not just how it looks or where the robot moves.
Real industrial dynamics
Train on production‑grade robots with accurate kinematics, calibrated dynamics, and stable control. Models are exposed to the same physics they will encounter in deployment.
Designed for deployment
With over 100,000 robots deployed globally, UR provides an industrial platform that supports long‑running experiments, repeatability, and a realistic path from lab to factory.
Anders Billesø Beck, VP of AI Robotics Products, Universal Robots.The AI Trainer gives AI teams the confidence to focus on building better models, knowing the data they gather will carry through to real industrial conditions. It also empowers industrial teams to capture the right use-case data and fine-tune models so they’re truly ready for deployment.
What our system enables
High‑fidelity task data
Capture real task demonstrations that include contact, compliance, and interaction, not just motion paths.
VLA model training
Generate datasets suitable for training and fine‑tuning Vision‑Language‑Action models for physical task execution.
Lab‑to‑factory continuity
Train on the same industrial robots you intend to deploy, reducing sim‑to‑real friction.
How the system works
Demonstration
- Human operators guide leader robots through real tasks
- Follower robots mirror motion in real time
- Demonstrations are repeatable and consistent
Control & interaction
- Ultra-low latency for precision tasks
- Force feedback to feel contact directly
- Good ergonomics for long-term operation
Sensing & data capture with Scale AI software
- Four synchronized camera streams
- Robot state, motion, and force data captured together
- Time‑aligned datasets across all signals
Data pipeline
- Runs on Universal Robots’ AI Accelerator
- Integrates with Scale AI's annotation pipelines and quality control
What’s included in the system
UR collaborative robots
Leader and follower robots for task demonstration and execution.
AI Accelerator
On‑robot AI compute using NVIDIA Jetson Orin, with integrated Orbbec 3D vision for high‑fidelity data capture.
Scale AI software
Recording software for automatically capturing and managing imitation-learning robot datasets for model training.
Vention platform
Industrial‑grade system, developed with Vention, for multi‑robot leader–follower setups.
Robotiq grippers
Production‑ready grippers for realistic grasping, handling, and contact‑rich demonstrations.
Recording data for model training with Scale AI software
The Scale AI software enables end users to seamlessly capture high-quality multimodal data for Physical AI development. During each demonstration collected on UR AI Trainer, the Scale software records synchronized motion, force, and visual data, producing structured datasets required to train Vision-Language-Action (VLA) models.
This allows for scalable data capture across live UR robot deployments and creates continuous feedback that accelerates the development and optimization of physical AI systems.
Leveraging UR AI Trainer and Scale’s software, Universal Robots and Scale AI will release a large-scale industrial dataset later this year.
Who this is for
AI and robotics research teams working on manipulation and Physical AI Advanced industrial R&D groups Technology partners developing AI models System integrators working on learning‑based automation
Get started with imitation learning
Complete the form to speak to an expert about your application, data strategy, and setup.