Intrinsic releases Intrinsic Core as open source under Apache 2.0, with a reference design for CNC machine tending

Intrinsic, a robotics software group inside Google, released Intrinsic Core on September 22, 2026, putting parts of its robot software platform on GitHub as open source under the Apache 2.0 licence. The company announced it at ROSCon 2026 in Toronto, together with the Open Machine Tending Solution, an open reference design for loading and unloading a CNC machine. The main repository held 6,074 files when we read it through the GitHub API on October 5, and GitHub counts 68.6% of its code as C++. We have not installed or run any of it. This piece reports what Intrinsic states in its announcement and what its two repositories contain.

Takeaway points

• Intrinsic released Intrinsic Core on September 22, 2026 under the Apache 2.0 licence, in a repository of 6,074 files that GitHub counts as 68.6% C++.
• The FoundationPose model weights used for pose estimation are not in the repository and are not covered by the Apache 2.0 licence, and Intrinsic’s setup guide says ARM processors are unsupported.
• Intrinsic has not said which reference design comes next or when, and we have not installed or run the software.

What is Intrinsic Core and what did Intrinsic release?

Intrinsic Core is, in the words of Intrinsic’s announcement, “an open source set of ROS-compatible capabilities” for building robot applications. ROS is the Robot Operating System. The announcement says these are the same capabilities and services Intrinsic uses for real manufacturing deployments, and that they run on local hardware.

The announcement lists seven things in the release. Each description below is Intrinsic’s claim, not something we measured.

The announcement does not name the supported robots, grippers and cameras.

Intrinsic said on February 25, 2026 that it had joined Google, where it works as “a distinct group within Google”. The same post says it was founded inside Alphabet in 2021. The code carries a different note. The readme files of both repositories state: “This is not an officially supported Google product.” Anyone who wants to contribute code must sign Google’s Contributor License Agreement, the contributing file says.

Two GitHub organisations carry the Intrinsic name. The announcement and the product page on Intrinsic’s site both link to the repository in the organisation named intrinsic-ai, which GitHub lists with 14 public repositories. The other, intrinsic-dev, is titled Intrinsic Developer Resources, has 2 public repositories, and neither is Intrinsic Core.

What is in the Intrinsic Core repository on GitHub?

The Intrinsic Core repository holds 6,074 files in nine module folders plus build, tutorial and third-party code, and its licence file is the Apache License, Version 2.0. GitHub records the repository as created on September 8, 2026, two weeks before the announcement. It has two tagged releases, 20260921.0 and 20260922.0.

The language shares are our arithmetic on the byte counts GitHub reports. Starlark is the language of the Bazel build files.

The readme names nine modules: a runtime, the control framework (which the code calls ICON), motion planning, perception, inference, a software development kit, interface definitions, hardware drivers and kinematics. It says the runtime is packaged as a k3s environment, a version of the Kubernetes container system, and gives ROS 2 Lyrical Luth as the ROS version.

The folder for hardware modules has entries named for ABB, KUKA and Universal Robots, next to others such as a loopback module and a Gazebo stub. The folder of robot definitions has entries for ABB, FANUC, KUKA and Universal Robots. We found no hardware module folder named for FANUC, and we did not check whether FANUC arms are driven another way.

On October 5 the repository listed 65 contributor accounts, one of them an automated dependency bot, and 65 forks.

The release of September 22 also carries three downloads that are not source code.

The readme points to a companion repository that plugs MoveIt 2 into Intrinsic’s motion and grasp planning. MoveIt is the software named in Qualcomm’s agreement to acquire PickNik Robotics.

What does the Intrinsic Core reference design for CNC machine tending contain?

The Open Machine Tending Solution is a Python application of 202 files that runs one cycle: pick a part, load the machine, wait for machining, unload the machine, and return the part. Its readme says the default cell is a UR5e arm with a CNC enclosure and a Schunk vise. A second configuration, for a lab cell with a UR3e arm, has no CNC machine.

The readme describes two ways of picking a part. One uses a 3D camera and pose estimation for parts placed at random. The other works from fixed slots on a tray, which the readme calls a blind grid pick. Digital outputs open and close the CNC machine door and work the vise clamp, and a separate skill drives a Robotiq Hand-E gripper.

Grasp planning in the reference design is simpler than the announcement suggests. The readme says the grasp planner defaults to a method called cuboid_center, which derives a top-down grasp from the pose estimate. Planning the grasp with MoveIt is optional, and the readme warns that the option will fail at runtime until the user has completed a separate integration.

Intrinsic’s announcement says the design can be customised for hardware from FANUC and Universal Robots. The file tree of the design’s repository has Universal Robots configuration files and no path containing the word fanuc. The FANUC robot definitions sit in the main repository.

The setup guide says the first build takes about 50 minutes, because it compiles the Gazebo simulator and other dependencies from source. The guide’s example deploys the cell in simulation, not on a real robot.

Intrinsic gives a market reason for choosing this job. Its announcement says that in the United States and Europe there are tens of thousands of businesses that fabricate parts and that only a small fraction have some form of automation. It gives no source for that count. It calls the design the first of a series.

Which parts of Intrinsic’s platform stay outside the Intrinsic Core release?

Intrinsic names three things as enterprise services, apart from the open release: “advanced AI models, Intrinsic Flowstate, and industrial-grade cloud services.” Intrinsic describes Flowstate as its web-based development environment and simulation engine. The announcement says applications built with the open code work with those services with no code rewrites. It does not give their prices or licence terms. The readme of the MoveIt companion repository refers to users with a Flowstate licence.

Three more limits are written in the repositories, not in the announcement.

The pose estimation model is not Intrinsic’s to open. The reference design’s readme says the FoundationPose weights are not included in the repository and are not covered by the Apache 2.0 licence. The build downloads them from Nvidia under the Nvidia Open Model License. “Intrinsic does not distribute these weights,” the readme says.

The name is not open either. The trademark file says: “The Apache 2.0 license does not grant permission to use Intrinsic trademarks or logos.” Anyone who distributes a modified version must give it a different name.

The reference design’s build also downloads prebuilt files from its release page. The readme lists six of them, and the release of September 22 carried five when we read it, 3.88 GB in total by our sum. They include gripper and camera driver bundles and a pretrained segmentation model. We did not check whether each can be rebuilt from the published source.

The scope is narrow by design. The Robot Report, a trade publication, reported that Intrinsic Core is a platform for robotic arms and does not support mobile robots or humanoids. That puts it apart from the humanoid stories of the same weeks, such as Boston Dynamics opening an application center for its Atlas humanoid at Hyundai’s Georgia plant and Agility Robotics’ merger filing with its Digit revenue and deployment figures.

What computer does Intrinsic Core need to run?

Intrinsic Core needs a PC with an x86-64 processor running Ubuntu, with 32 GiB of memory at the least, according to the setup guide in the repository. The guide says: “ARM architectures are strictly unsupported at this time.” It adds that real-time control of robot hardware needs an Intel processor, while simulation can also run on AMD. The table lists the guide’s recommended specification.

Intrinsic’s own files disagree on two points.

We do not know which Ubuntu statement is meant to stand. Intrinsic’s announcement mentions none of these requirements. It says only that using the software “continues to require some basic proficiency in robotics”.

What has Intrinsic not published about Intrinsic Core, and have we tested it?

We have not tested Intrinsic Core. We have no robot arm and no CNC machine, and we did not install the software or run the simulated cell. Every figure here comes from Intrinsic’s announcement, its product page, the files in its two repositories and the GitHub API, all read on October 5, 2026. The star, fork and contributor counts will have moved since.

Several things are not in what Intrinsic has published. The announcement gives no control rate, cycle time or grasp success figure for the software or for the machine tending cell. It does not say which reference design follows this one, or when. It does not list the supported robots, grippers and cameras, which a reader has to work out from the folders. It does not say why the code states that it is not an officially supported Google product seven months after Intrinsic joined Google. And it does not reconcile the two Ubuntu requirements in its own files.