Google Open-Sources the Plumbing of Industrial Robotics: A Close Look at Intrinsic Core

Most talk about “physical AI” is about robot brains: foundation models that look at a cluttered bench and decide what to pick up. Far less gets said about the dull layer underneath, which is the software that turns “pick up that part” into joint torques a thousand times a second without hitting the fixture. That layer is also where most industrial robotics projects burn their budgets. On 22 September 2026, at ROSCon 2026 in Toronto, Alphabet’s robotics software unit Intrinsic open-sourced a large chunk of it. The project is called Intrinsic Core, and it ships under the Apache 2.0 licence.

What was released

According to Intrinsic’s announcement and the public repository at github.com/intrinsic-ai/intrinsic-core, Intrinsic Core bundles these pieces:

Intrinsic Control: a hardware-agnostic, real-time control framework (the README calls the motion engine ICON). It can adjust a robot’s trajectory mid-move based on sensor feedback.
Motion planning: generation of collision-free paths.
Grasp planning: adapts the gripper’s approach to how an object is actually lying, with sensor feedback.
Pose estimation: 6-DoF object detection built on NVIDIA’s FoundationPose model, so parts can be located without rigid mechanical fixtures.
Simulation and digital twin: services built on Gazebo, plus automated camera calibration.
Intrinsic-ROS drivers: preconfigured interfaces for robots, grippers and cameras.
A local runtime and SDK: a preconfigured, containerised environment (k3s, a lightweight Kubernetes, according to the README) that runs on local hardware rather than in a cloud.

The codebase is a mix of C++, Python and Go, built with Bazel. The README lists Ubuntu 22.04, 24.04 and 26.04 LTS, and interoperability with ROS 2 “Lyrical Luth”. Its top-level modules (intrinsic_control, intrinsic_motion_planning, intrinsic_perception, intrinsic_inference, intrinsic_kinematics, intrinsic_hardware, intrinsic_sdk) read like a map of what robotics integrators usually have to glue together themselves. The repository is brand new: when I checked it the day after launch it had only about 130 stars and a few dozen commits, so there is no community track record yet.

Intrinsic also published the Open Machine Tending Solution (OMTS), an open reference design for CNC machine tending. That means a robot loading raw stock into a milling machine and unloading finished parts. OMTS runs on Intrinsic Core and the Open Robotics Suite, and can be adapted to arms from FANUC and Universal Robots.

What stays closed

The line is drawn deliberately. Intrinsic’s own post names what remains proprietary: its AI models (branded Intrinsic Intelligence and the Intrinsic Vision Model), Flowstate, its enterprise orchestration and development platform, plus “industrial-grade” cloud services and enterprise support. Intrinsic says applications built on Core can move onto its commercial products without rewrites. That is the key to the business model.

Why this is the Android play

The framing is not subtle. When Google folded Intrinsic out of Alphabet’s “Other Bets” and into Google proper in February 2026, CNBC reported the company wants it to be the “Android of robotics”, working more closely with Google DeepMind and using Gemini models. Android won by giving away the operating system and monetising services and distribution on top. Intrinsic Core applies the same logic to factories: make the control and planning layer free and ubiquitous, then sell the intelligence and fleet management that plug into it.

The ROS connection makes the strategy credible. In December 2022 Intrinsic acquired the Open Source Robotics Corporation, the commercial arm of Open Robotics and home of many core ROS and Gazebo engineers. ROS, Gazebo and Open-RMF themselves stayed with the non-profit Open Source Robotics Foundation. Launching Core at ROSCon, with ROS 2 drivers and Gazebo simulation built in, is a bid to become the default production-grade layer for the millions of developers who already prototype on ROS.

Why it matters

For developers, the value is mostly in time saved. Real-time control that tolerates sensor-driven corrections, collision-aware planning and reliable 6-DoF perception are each hard engineering problems. Today they are usually solved per project, per vendor, often with proprietary controller languages. A permissively licensed stack that spans FANUC and UR hardware lowers the barrier for small integrators and machine shops. SiliconANGLE, citing Intrinsic, notes that only about 8% of US and European machine shops use automation. That figure is Intrinsic’s and I have not independently verified it.

It also shifts the competition. NVIDIA is pushing Isaac and its own robot foundation models, and Tesla and Amazon build vertically integrated stacks. An Apache-licensed alternative backed by Google puts pressure on anyone charging for the plumbing.

Open questions

Governance. This is a corporate open-source project, not a foundation project. Roadmap control, contribution policy and long-term commitment all sit with Google. Google has open-sourced things before and later let them go quiet.
Real-time claims. “Hardware-agnostic real-time control” is easy to say and hard to prove. Deterministic latency depends on kernels, fieldbuses and controller firmware. Independent benchmarks on real cells don’t exist yet.
Dependency gravity. Pose estimation leans on NVIDIA’s FoundationPose, and the best models remain proprietary. “Open core” may in practice mean open until you need the parts that make it good.
Overlap with ROS 2. Whether Core complements ROS 2 packages like MoveIt or quietly competes with them will decide how the ROS community receives it.

For now, the most useful thing a robotics developer can do is clone the repo, run OMTS in simulation, and see how much of their own glue code it replaces.