Alphabet open-sources its entire robotics stack: Intrinsic Core under Apache 2.0 at ROSCon 2026

Alphabet open-sources its entire robotics stack: Intrinsic Core under Apache 2.0 at ROSCon 2026

🔎 Alphabet just released "the Android of robotics"

On September 22, 2026, at ROSCon 2026 (Toronto), Intrinsic Innovation — Alphabet's robotics unit — announced Intrinsic Core, its foundational infrastructure for intelligent robots, under the Apache 2.0 license. A preconfigured, ROS-compatible robotics environment designed to run on local hardware (SiliconANGLE).

Forget the idea of a showcase project. This is not a tutorial framework: it's the production stack used in real industrial deployments (ByteIota). Real-time control, motion planning, grasp planning, pose estimation, simulation, calibration, drivers — building blocks that integrators previously had to buy or rebuild are going open source, overnight.

Forbes summed up the move with a phrase already making the rounds in the robotics world: "the Android of robotics." The analogy is precise. Open-source the infrastructure layer, an ecosystem builds on top of it, paid cloud services sit above. Google has already played this tune, with historic success.

And the timing is no accident: NVIDIA unveiled Isaac ROS 5.0 at the same conference. Open physical AI now has two serious foundations — and, importantly, they complement each other more than they compete.

Key Takeaways

• Announcement: Intrinsic Core released on September 22, 2026 at ROSCon 2026 (Toronto) by Intrinsic Innovation, Alphabet's robotics unit, under the Apache 2.0 license.
• Contents: Intrinsic Control (hardware-agnostic real-time control), pose estimation via NVIDIA FoundationPose, motion planning and grasp planning, Gazebo simulation services, automated camera calibration, Intrinsic-ROS drivers, asset models, and digital twin.
• Reference design: Open Machine Tending Solution (OMTS), ready to use for machine tending, with support for FANUC and Universal Robots arms.
• Access: GitHub repos intrinsic-ai/intrinsic-core and intrinsic-ai/intrinsic-omts, commercial use without royalties or permission.
• Limitation: advanced AI models, Intrinsic Flowstate, and the industrial cloud remain paid.
• Context: Intrinsic, founded in 2021, was integrated into Google's operations in February 2026 to accelerate physical AI.

Recommended Tools

Intrinsic Core: what exactly is it?

Direct answer: a complete, preconfigured robotics environment — ROS-compatible, running on local hardware — that covers the physical AI pipeline end to end. Not a library, not an SDK you have to assemble yourself.

Intrinsic Control: real-time control, once the privilege of major industrial players

The central component is called Intrinsic Control. It's a hardware-agnostic real-time control framework, capable — and this is the detail everyone at the conference is talking about — of dynamically adjusting a robot's trajectory mid-motion, based on sensor feedback (SiliconANGLE).

In concrete terms: the robot no longer follows a fixed trajectory, it corrects on the fly. This level of responsiveness has until now been locked away inside proprietary stacks. It becomes an open source component that any integrator can deploy.

The perception-planning pipeline, delivered functional

The rest of the release fits in a list that reads like the table of contents of a robotics curriculum — except it's functional and preconfigured (The AI Insider):

• Pose estimation: position and orientation of 3D parts in 6 degrees of freedom, via NVIDIA FoundationPose.
• Motion planning: collision-free trajectory generation, without joint-by-joint programming.
• Grasp planning: the gripper adapts to the object's actual position thanks to the sensors.
• Simulation: Gazebo services to test everything before touching the hardware.
• Automated camera calibration and preconfigured ROS drivers for sensors, third-party grippers, and 3D cameras.

Asset models and digital twin are included. One last point often left unsaid: Intrinsic Core applications also work with Intrinsic Flowstate, Intrinsic's AI models, and cloud services, without any rewriting. That's exactly where the business model lies — more on that later.

Why this move is major: the end of proprietary building blocks

Direct answer: because control, motion planning, grasp planning, simulation, pose estimation, and camera calibration — historically proprietary — are going open source for free, under a license that allows commercial use without royalties or permission.

It's Forbes that set the analytical framework: Google is offering "the Android of robotics." The infrastructure layer becomes a common good, the ecosystem settles on top of it, and monetization shifts to services. The exact scope deserves a table:

Intrinsic's trajectory sheds light on the strategy. Founded in 2021 as an Alphabet "Other Bet," the company acquired Open Source Robotics Corporation in December 2022 — the commercial arm of the organization behind ROS — making Brian Gerkey its CTO. After a 20% workforce reduction in January 2023, Intrinsic left the fold of Alphabet's long-term bets in February 2026 to be integrated into Google, with its CEO Wendy Tan White reporting to Hiroshi Lockheimer — and direct connections to the Gemini models, Google Cloud, and Google DeepMind (Forbes).

My take: this isn't philanthropy, it's ecosystem warfare. Google learned with Android that an OS can be free when the services on top are not. By open-sourcing the layer everyone dreaded having to rebuild, Intrinsic ensures that the next generation of robotics applications will be born on its foundations.

A final credibility signal, and not the least: Brian Gerkey (Intrinsic's CTO and co-creator of ROS) and Geoffrey Biggs (CTO of Open Robotics) emphasized that the shared code is the heart of their own stack (Digest AI). You don't release a prototype: you release what powers real industrial deployments.

Open Machine Tending Solution: the stack translated to the shop floor

Direct answer: the OMTS is what keeps Intrinsic Core from being a beautiful GitHub repository that nobody uses. It's the proof by use case.

The Open Machine Tending Solution is a ready-to-use reference design for a specific shop floor scenario: machine tending — loading and unloading machine tools with a robotic arm. It supports FANUC and Universal Robots arms, two industry standards (Digest AI).

The covered cycle is complete: workcell setup, simulation, motion planning with MoveIt 2, deployment on real hardware (ByteIota). In other words, the entire path from "I have an arm in a box" to "it's working in production."

Why start with machine tending? Because it's the ideal use case for an open stack: repetitive, well-defined, easy-to-calculate ROI. It's the "hello world" of industrial robotics — and the first place where an integrator will measure the difference between a free foundation and a license with six zeros.

Intrinsic Core vs. NVIDIA Isaac ROS 5.0: complementary, not competing

Direct answer: the two stacks don't replace each other, they complement each other — and the fact that they're launching at the same conference says a lot about the state of open physical AI.

NVIDIA announced Isaac ROS 5.0 at ROSCon 2026: AI agent skills for ROS, FoundationStereo, tracking via FoundationPose, migration to ROS 2 Lyrical (ByteIota). The division of labor is clear:

Intrinsic Core brings the integrated, hardware-agnostic runtime; Isaac ROS brings GPU inference on NVIDIA hardware. Note the loop closing: Intrinsic's pose estimation relies precisely on FoundationPose, a NVIDIA model.

And NVIDIA isn't stopping at ROS: with Hugging Face, the company is already pushing LeRobot, Isaac GR00T 1.7, Isaac Teleop and Cosmos 3 open source for robotics. Physical AI is being commoditized from both ends — infrastructure on Google's side, models on NVIDIA's side.

For a robotics team, the real question is therefore not "Intrinsic or NVIDIA," but "which combination." If your shop is already NVIDIA-centric, Isaac ROS is the natural choice. If you want a runtime that doesn't tie you to any chipmaker, Intrinsic Core is the most agnostic option on the market — and nothing prevents you from mixing the two.

Getting Started with Intrinsic Core: the shortest path

Direct answer: two public GitHub repos, a preconfigured local runtime, and a golden rule — start with simulation before touching any arm.

The repos are immediately accessible:

git clone https://github.com/intrinsic-ai/intrinsic-core.git
git clone https://github.com/intrinsic-ai/intrinsic-omts.git

My first piece of advice: don't hook up hardware on day one. The built-in Gazebo simulation services let you build, test, and virtually break your workcell — that's exactly why the pipeline puts simulation before deployment. If you have a FANUC or Universal Robots arm on hand, OMTS then gives you the signposted path all the way to real deployment.

Second tip: if you already know ROS 2, the learning curve will be gentle. Intrinsic Core is ROS-compatible, with preconfigured drivers for sensors, third-party grippers, and 3D cameras. Your MoveIt 2 and Gazebo skills remain assets, not liabilities.

On the intelligence side, be clear about the scope: Intrinsic Core is infrastructure, not a brain. Advanced AI models are Intrinsic's paid layer. The choice is yours: plug in cutting-edge cloud models — GPT-5.5 currently dominates agentic benchmarks (98.2, June 2025 ranking), with Gemini 3.1 Pro holding strong in general use (92) — or run open source models locally, which our guide on open source AI agents with Ollama walks through step by step.

Finally, to document your experiments — technical docs, blog, video demos — a small VPS is more than enough. Hostinger starts at around €5/month (September 2026, check hostinger.com), and that's plenty for publishing your hands-on findings.

Apache 2.0, MIT, open weights: Intrinsic rides a groundswell

Direct answer: Intrinsic Core is not an isolated case — it's the confirmation of a movement: the major players are opening up their foundations to win the ecosystem, and monetizing what sits on top.

Look at the recent sequence. DeepSeek V3.1 moved to the MIT license, confirming the rise of Chinese open source in language models. Poolside released Laguna M.1, a 225B model under Apache 2.0 designed for coding agents. MiniMax M3 challenges GPT-5.5 with 1M context in open weights. And now Alphabet is opening up its robotics infrastructure.

The common thread jumps out: permissive licenses (Apache 2.0, MIT) that allow commercial use without royalties, a foundational layer given away for free, and value shifting toward services. What changes with Intrinsic Core is the domain. Here, the foundation is not a language model: it's real-time control, trajectory planning, and simulation — building blocks that act on the physical world, with everything that implies in terms of safety, certification, and liability.

My prediction: physical AI has just crossed its Android moment. In eighteen months, the question will no longer be "which proprietary robotics stack to choose," but "which application to build on a foundation that has become common." Competitive advantage shifts one level higher up the stack.

❌ Common Mistakes

Mistake 1: Believing that "all of Intrinsic" becomes free

This is the point Forbes hammers home: control, the digital twin, and the drivers are free; advanced AI models, Flowstate, and the industrial cloud are not. Map out the boundary before setting a budget — the open source core does not cover the entire stack, far from it.

Mistake 2: Treating Intrinsic Core as a learning framework

It is not a tutorial framework, but the production stack used in real industrial deployments. Expect workshop-level complexity: workcell setup, calibration, real-time constraints. The upside is clear: what you learn is directly deployable.

Mistake 3: Underestimating the ROS prerequisite

ROS-compatible does not mean ROS-free. Intrinsic Core is preconfigured, but the ecosystem remains ROS 2: drivers, MoveIt 2, Gazebo. Check the compatibility of your hardware — sensors, third-party grippers, 3D cameras — via the Intrinsic-ROS drivers before making any purchase.

Mistake 4: Looking for an AI brain in the core

Pose estimation relies on NVIDIA's FoundationPose, but advanced AI models remain Intrinsic's paid layer. Plan your intelligence stack separately: cutting-edge cloud models for maximum performance, local open source models for control and cost savings.

❓ Frequently Asked Questions

Is Intrinsic Core really free for commercial use?

Yes. The Apache 2.0 license allows commercial use without royalties or permission, and the runtime is designed to run locally on your hardware. You only pay if you adopt the services sitting above the core: Flowstate, advanced AI models, or Intrinsic's industrial cloud.

Do you need NVIDIA hardware to use it?

No. Intrinsic Control is hardware-agnostic and the environment is preconfigured for local hardware. That said, pose estimation relies on NVIDIA FoundationPose, and Isaac ROS 5.0's GPU inference assumes NVIDIA hardware. The two stacks combine, they don't replace each other.

What's the difference with ROS 2?

ROS 2 is a middleware; Intrinsic Core is a preconfigured environment built on top of it. It integrates real-time control, motion and grasp planning, simulation, calibration, and drivers, while relying on existing ROS 2 tools — the OMTS explicitly uses MoveIt 2 and Gazebo.

Why is Google open-sourcing its robotics stack?

To replicate the Android model: open up the infrastructure layer to win the ecosystem, monetize the services above it. Context sheds light on the decision: Intrinsic was integrated into Google in February 2026 to accelerate physical AI, connected to Gemini models, Google Cloud, and Google DeepMind.

Can you test without a physical robot?

Yes. The Gazebo simulation services let you test everything before touching hardware, and the Open Machine Tending Solution explicitly covers the cycle of workcell setup, simulation, then deployment. A PC capable of running the preconfigured local runtime is enough to get seriously started.

✅ Conclusion

Alphabet has just moved real-time control, planning, and robotic simulation from the status of proprietary advantage to that of a common good under Apache 2.0 — and the business will now be played one layer above. The intrinsic-ai/intrinsic-core repository is public: clone, simulate, build.