Google’s Intrinsic open-sources Intrinsic Core under Apache 2.0

Google’s Intrinsic announced Intrinsic Core, an open-source set of ROS-compatible robotics capabilities under the Apache 2.0 license, plus an Open Machine Tending Solution reference design, at ROSCon 2026 in Toronto.

ROBOTICS desk — Google’s industrial robotics software unit is putting the boring-but-critical stack (control, grasp, sim) on GitHub so shops stop reinventing Physical AI plumbing. Physical AI means software that has to run on a machine in the real world, not only answer a question on a screen.

Intrinsic published an open-source release it calls Intrinsic Core. The record is the company blog, “Introducing Intrinsic Core™,” dated September 22, 2026. The post says the announcement is at ROSCon 2026 in Toronto. ROSCon is the conference for people who build robot software on ROS. ROS, short for Robot Operating System, is the common open toolkit a robot uses to talk to its arm, its cameras, and its other parts. These lines are Intrinsic’s. This desk did not attend the conference.

What the license allows. The blog says the release is under the Apache 2.0 license. Apache 2.0 is a common open-source license: you may use, change, and ship the code, including in a product you sell, if you keep the license and copyright notices. The public repository’s own page labels the project Apache-2.0, and the README says the same. File the license as Intrinsic’s. This desk did not install the stack.

What Intrinsic Core is, in the blog’s words. It is an open-source set of ROS-compatible capabilities for building robotic applications, aimed at what the post calls physical AI. Physical AI, here, means software that has to work on a machine in the real world, not only answer a question on a screen. The post says the pieces are the same kind of capabilities Intrinsic uses for real manufacturing deployments, and that they are meant to be combined instead of coded from scratch. It also says Intrinsic Core works with the Open Robotics Suite. That suite is the blog’s name. This filing does not invent a second definition for it. A hardware-agnostic real-time control framework and a digital twin are included. Hardware-agnostic means the control software is meant to run on more than one brand of arm. Real-time means the control loop has to answer on a clock. A digital twin is a software copy of the workcell, the robot and the machines around it, used to try the setup before or while you build. File that frame as the blog’s.

The building blocks the blog lists. Intrinsic Control is the hardware-agnostic real-time control framework. The post says it uses sensors so the robot can change what it is doing in the middle of a move, and that you can swap arms, grippers, and sensors without rewriting the drivers. A gripper is the hand that picks the part. A driver is the software that speaks to one piece of hardware. Pose estimation uses NVIDIA FoundationPose. Pose estimation means figuring out where a part is and which way it is turned. The blog calls that a 6-DoF estimate. 6-DoF means six degrees of freedom: three for position and three for rotation. The stated point is to pick parts up without locking them into expensive fixtures. Motion planning, the post says, generates paths that avoid collisions, so you do not program every joint by hand. Grasp planning lets grippers adapt to where a part sits, using sensor feedback. Simulation services run on Gazebo, a robot simulator, so you can test the cell as you build. Camera calibration is the setup step that lines the camera up with the robot. Intrinsic-ROS drivers are pre-configured drivers for supported robots, grippers, and 3D cameras. File that list as the blog’s. This desk did not score any of it.

The reference design shipped with it. The blog also releases the Open Machine Tending Solution, shortened to OMTS. Machine tending means a robot loads and unloads a machine, such as a mill, so a person is not feeding every part. The post calls OMTS a ready-made reference design for a machine-shop job. It says the design runs on Intrinsic Core and on the Open Robotics Suite, and that you can customize it for hardware from FANUC and Universal Robots, and for foundation models such as NVIDIA FoundationPose. A reference design is a working example you start from. It is not a report that a named shop is already running it in production. The README on the Intrinsic Core repository points that design at a second public repository, intrinsic-omts. File the OMTS description as the blog’s, and the second repository as the README’s. This desk did not run either one.

Two shop counts, and they are not the same sentence. The blog says that in the U.S. and Europe alone there are tens of thousands of businesses that fabricate parts, and that only a small fraction have some form of automation. The product page says hundreds of thousands of machine and fabrication shops around the world need better access to affordable adaptive robotic solutions. Tens of thousands, limited to the U.S. and Europe, is the blog’s. Hundreds of thousands, worldwide, is the product page’s. This filing does not merge them into one number, and it does not treat either line as a census this desk counted.

What this release does not replace. The blog says solutions built with Intrinsic Core automatically work with other Intrinsic offerings, including enterprise services: advanced AI models, Intrinsic Flowstate, and industrial-grade cloud services, with no refactoring required. Flowstate is Intrinsic’s commercial application builder. The company’s February 2026 post describes it as a web-based development environment. Working with those services is not the same as giving them away. This open release is core parts of the platform. It is not a claim that every commercial Intrinsic service is now free, and it is not a claim that Intrinsic Core replaces them. The repository README adds a separate limit: “This is not an officially supported Google product.” File both limits.

Why the headline says Google’s Intrinsic. Intrinsic’s own post of 25 Feb 2026 says the company had joined Google. It says Intrinsic was founded as an Alphabet Other Bet in 2021, and that it would continue as a distinct group within Google. That is prior corporate status. The 22 Sep open-source post does not restate the join. File “Google’s Intrinsic” as that February record. Do not read this September release as a new Google product announcement, and do not skip the README line that it is not an officially supported Google product.

Do not invent a scoreboard. The blog and the product page do not print a benchmark, a timing number, or a grasp-success rate for Intrinsic Core. This filing does not add one. Do not print a star count on the code host. Stars change, and this card does not use them. A public repository is not a test this desk ran, and it is not a promise that a machine shop can drop the code in and skip integration.

Plain English for the rest of the card: Intrinsic Core = the open-source robotics building blocks in this release. ROS = Robot Operating System, the common open toolkit for robot software. Apache 2.0 = an open-source license that lets you use and ship the code if you keep the notices. physical AI = software that has to work on a machine in the real world. hardware-agnostic = meant to run on more than one brand of robot. real-time = the control loop answers on a clock. digital twin = a software copy of the workcell. gripper = the robot hand. driver = the software that talks to one device. pose estimation = where a part is, and which way it faces. 6-DoF = three directions of position plus three of rotation. Gazebo = the robot simulator named for the simulation services. machine tending = a robot loading and unloading a machine. OMTS = Open Machine Tending Solution, the reference design. Flowstate = Intrinsic’s commercial builder for robot applications. This filing is the 22 Sep open-source release. It is not a benchmark and not a shutdown of the paid services.

PRIMARY here: Intrinsic’s 22 Sep 2026 blog post, “Introducing Intrinsic Core™” — Tier A PRIMARY, the company’s own announcement. The Intrinsic Core product page and the public repository github.com/intrinsic-ai/intrinsic-core are the company’s own same-day records, not a second newsroom. The ROSCon announcement, the Apache 2.0 release, the ROS-compatible capability list, the hardware-agnostic real-time control line, the digital twin, FoundationPose pose estimation, motion planning, grasp planning, Gazebo simulation, camera calibration, the Intrinsic-ROS drivers, the OMTS reference design, the FANUC and Universal Robots customization line, and the U.S. and Europe “tens of thousands” line are the blog’s. The “hundreds of thousands” worldwide shop line and the partner names on that page are the product page’s. The Apache-2.0 label, the second repository for OMTS, and the “not an officially supported Google product” line are the repository’s. The join-Google history is the 25 Feb 2026 Intrinsic post. NOT claimed: a benchmark score, a grasp-success rate, a star count, that this replaces all commercial Intrinsic enterprise services, that Flowstate or the cloud services are now free, that this desk installed the stack or attended ROSCon, a stock tip, or investment advice. Distinct from the already-filed boston-dynamics-rmac, einride-nvidia-hyperion, and meshy-mora.