Universal Robots: building the AI Accelerator for cobots on NVIDIA Isaac and Jetson

Collaborative robot arms are usually programmed point by point, which suits repeated tasks but breaks down when parts arrive in varying positions or production changes often. NVIDIA's account of the project frames the problem as high-mix, low-volume work where a cobot has to perceive its surroundings and adapt, against a background of labor shortages.

For Universal Robots the question was a product one: how to give developers and integrators a supported, repeatable way to add AI perception and planning to its installed base of arms instead of each partner assembling its own compute, camera and software stack.¹²

2024: collaboration and first demo

In March 2024 Teradyne Robotics, the group that includes Universal Robots and Mobile Industrial Robots (MiR), announced work with NVIDIA. UR showed a UR5e running path planning with NVIDIA cuMotion from Isaac Manipulator on a Jetson AGX Orin module under its new PolyScope X software. The same week MiR launched the MiR1200 Pallet Jack, whose pallet detection runs on Jetson AGX Orin and learned from a training set of over 1.2 million images, both real and synthetic; MiR states that the robot's design targets ISO 3691-4.

October 2024: AI Accelerator launch

UR introduced the AI Accelerator in Odense in October 2024 as a hardware and software kit for commercial and research developers. Isaac software executes on the Jetson AGX Orin module, which UR pairs with an Orbbec Gemini 335Lg 3D camera, and ships demo programs for pose estimation, tracking, object detection, path planning, image classification, quality inspection and state detection. UR's product page lists bundles with or without a control box and an e-Series arm.

2025: version 1.1 and partner applications

For GTC in March 2025 UR released version 1.1, which integrates Isaac ROS, improves pose estimation and path planning through Isaac Manipulator, and moves the camera link from USB3 to GMSL. Five partners (3D Infotech, T-Robotics, AICA, Acumino and Groundlight) demonstrated inspection, language-driven machine tending, reinforcement learning assembly, two-arm assembly and workpiece detection. NVIDIA's case study adds that UR uses cuMotion, FoundationPose and FoundationStereo, Isaac ROS with NITROS, and Isaac Lab with IndustReal algorithms, and that the AI Accelerator spans the e-Series and UR Series arms including the UR15.

Sim-to-real assembly

According to NVIDIA, UR trained reinforcement learning policies for gear grasping, transport and insertion entirely in simulation with Isaac Lab, using domain randomization and synthetic data, and the policies assemble gears from random start poses without fine-tuning on real hardware.¹²³⁴⁵⁶

What others can learn: a robot maker can package edge compute, a camera and accelerated libraries as one supported kit, so integrators build applications rather than plumbing, and partner demos can prove the kit on concrete tasks before customers buy. Training contact-rich skills in simulation with randomized conditions is worth testing for tasks that are slow to teach by hand. Limits: the speed-up figures (50 to 80 times in 2024, up to 100 times later) have no published baseline, the sources name partner demonstrations but no end-user production site, and the 2024 MiR safety statement is a design intent, not a certificate shown in the sources. Fixed, well-presented parts are often handled with teach-pendant programming or simple 2D vision without a GPU module.