Tools

Alphabet's Intrinsic Open-Sources Its Robot Arm Software Stack Under Apache 2.0

September 23, 2026 · 15:08 UTC · Tools

TL;DR

Intrinsic Core is now on GitHub under Apache 2.0. Intrinsic, the Alphabet robotics group that moved into Google in February, announced it on September 22 at ROSCon 2026 in Toronto. It is the software layer Intrinsic uses for its own manufacturing deployments: a real-time control framework, motion and grasp planning, 6-DoF perception built around NVIDIA's FoundationPose, a Gazebo digital twin, camera calibration and hardware drivers, all packaged as a single-node Kubernetes runtime that runs on a workstation. It ships with an open reference design for CNC machine tending on FANUC and Universal Robots arms. The paid layer (advanced AI models, Flowstate, cloud services) stays closed.

What actually shipped

This is not a single library dumped over the wall. The main repo is about 6,000 files and roughly 700,000 lines of C++, Python and Go, organized into modules that each map to a job a robot cell needs done. The first tagged release, 20260922.0, landed on September 22 with three artifacts: a base runtime tarball, the inctl control CLI, and the source bundle.

intrinsic_runtime: a preconfigured execution engine packaged as a k3s cluster. It handles process lifecycle, scheduling and state sync on top of ROS 2.
intrinsic_control (ICON): the real-time motion engine. The README says it can switch controllers within a single control cycle based on live sensor feedback, behind a hardware abstraction layer meant to let you swap arms and grippers without rewriting drivers.
intrinsic_motion_planning: collision-free path generation across Cartesian and joint-space moves, with blending of multi-segment paths that have different velocity and acceleration profiles.
intrinsic_perception and intrinsic_inference: camera and point-cloud interfaces plus a local model server, with built-in support for NVIDIA FoundationPose for 6-DoF pose estimation of un-fixtured parts.
Simulation and calibration: a Gazebo-backed digital twin and automated camera-to-robot calibration.
intrinsic_sdk, intrinsic_kinematics, intrinsic_hardware: the extension points, solvers and device drivers.

Around it sit satellite repos: intrinsic-moveit for MoveIt 2 integration, camera drivers, a shared-memory IPC transport, and malloc-guard, a small C++ library that catches heap allocations inside real-time threads. That last one is the kind of tool you only write after a robot has stuttered mid-trajectory because somebody called new in the wrong loop.

The open release covers the whole path from camera frame to motor command, not just one planner.

Why Intrinsic is giving this away

The blog post is candid about the business model. Solutions built on Core "automatically work with other Intrinsic offerings, including our enterprise services which include advanced AI models, Intrinsic Flowstate, and industrial-grade cloud services." The free part is the plumbing. The paid part is everything you want once the plumbing works in production.

That is the Android playbook applied to robot cells: give away the layer everybody has to rebuild, and charge for what sits on top. Intrinsic CTO Brian Gerkey, who helped launch ROS itself, framed it to The Robot Report as a real commitment: "This is not us coming by and dropping off some code and saying, 'Have fun.' What we're sharing this week is the core of our own software stack."

The plumbing is free; the models and cloud on top of it are the product.

The reference design: a robot that feeds a CNC machine

The companion repo, intrinsic-omts (Open Machine Tending Solution), is an Apache 2.0 template for loading and unloading a CNC machine with an arm from FANUC or Universal Robots. Intrinsic's pitch is aimed at small fabrication shops: "In the U.S. and Europe alone, there are tens of thousands of businesses that fabricate parts, and only a small fraction have some form of automation." The company calls OMTS the first of a series and names metal bending and part inspection as other jobs these shops need help with.

Gerkey's argument for shipping a full solution instead of just components: "It's important to have those well-lit paths that show you this is not just a box of parts with some potential."

What it takes to run it

This is not a Raspberry Pi weekend project. The getting started guide asks for Ubuntu 26.04, x86-64 only, and an Intel CPU if you want real-time control of physical hardware (AMD is fine for simulation, ARM is "strictly unsupported"). Perception needs a dedicated NVIDIA card, RTX 3060/4060 class or better. The build uses Bazel, and the install flow is: set up k3s with a provided script, download the base runtime tarball, install inctl, optionally run a script that wires the GPU into k3s, then build OMTS.

16 GiB works only if you limit build jobs and stop the solution before rebuilding.

A homelab angle worth noting: the recommended network setup is two to three gigabit ports, one for LAN, one dedicated to the real-time robot controller, and an optional third for a PoE camera switch. If you already run k3s at home, the runtime model will feel familiar. The deploy step even reports its own health with refreshing honesty: "All pods are running; maybe they're working, maybe they're not!"

The repo is written for coding agents too

For AI builders, the most telling detail is in the repo root. Intrinsic ships an AGENTS.md that explains the platform's four primitives (assets, skills, the world model, and a behavior-tree executive) and a folder of nine agent skills covering Bazel, skill authoring, service authoring, robot motion and debugging. Its first hard rule is that skills are stateless: the runtime builds a new skill instance for every call, so loading model weights in __init__() will reload them on every invocation.

That is a direct acknowledgment that a lot of the code written against this stack will be drafted by Claude Code or Codex. Robotics has always had a steep ramp: real-time constraints, kinematics, calibration, drivers. Documentation shaped for agents lowers the ramp for the ROS developer who knows manipulation but not Bazel.

Caveats before you build a business on it

Arms only. Per The Robot Report, Core targets robotic arms and does not support mobile robots or humanoids. The ideal user already knows ROS and is aiming at commercial manipulation work.
Check the perception license. Intrinsic's own code is Apache 2.0, but FoundationPose comes from NVIDIA, and the upstream NVlabs repository license limits use to non-commercial research or evaluation. Confirm the terms of the model files you actually deploy before putting vision in a paying customer's shop.
Google CLA. Contributions require signing Google's Contributor License Agreement, and the README says it is "not an officially supported Google product."
Early days. The public repo was created on September 8 and had 239 stars at the time of writing. Expect the rough edges of a first release.

Key Takeaways

Intrinsic open-sourced its production robot arm stack, Intrinsic Core, under Apache 2.0 at ROSCon 2026 on September 22.
It covers control, motion and grasp planning, FoundationPose-based perception, a Gazebo twin and drivers, packaged as a local k3s runtime on ROS 2.
An open CNC machine tending reference design for FANUC and Universal Robots arms is the first of a planned series.
Enterprise AI models, Flowstate and cloud services stay proprietary; Core is the on-ramp to them.
Plan for Ubuntu 26.04, x86-64, 32 GiB+ RAM and an RTX-class GPU for vision, and verify FoundationPose licensing for commercial use.
The repo ships AGENTS.md and nine agent skills, so it is built to be extended with coding agents.

Sources: Intrinsic blog, intrinsic-core on GitHub, Intrinsic Core getting started guide, intrinsic-omts on GitHub, The Robot Report, SiliconANGLE, NVlabs FoundationPose license