Google’s Intrinsic is open sourcing the foundation of its industrial robotics platform today. The company is releasing control, motion planning, grasp planning, simulation, pose estimation and camera calibration software that was previously locked inside a proprietary product and making it open source under a permissive Apache 2.0 license.

It’s called Intrinsic Core. It’s ROS-compatible by design, so it works with the open source middleware that most of the robotics world runs on, and according to Intrinsic, these are the same capabilities used daily for real deployments. In other words, it’s not some stripped-down demo build. In addition Intrinsic is releasing what it calls the Open Machine Tending Solution, a reference design for AI-enabled CNC machine tending that runs on Intrinsic Core and supports models from robot makers like FANUC and Universal Robots out of the box.

The pitch is simple: robotics developers keep rebuilding the same plumbing. Drivers, kinematics, collision-aware path generation, calibration routines and so on. That means teams burn weeks if not months of development time on infrastructure before building something unique and actually starting to add value.

Intrinsic wants to hand that layer over for free. The more interesting question, of course, is what Google gets in return for this.

The Android comparison is Google’s

Earlier this year in February, Intrinsic stopped being an Alphabet “Other Bet” and folded into Google proper as a distinct group, with CEO Wendy Tan White reporting to Hiroshi Lockheimer. The move plugged Intrinsic into Gemini models, Google Cloud infrastructure and Google DeepMind research.

Tan White told CNBC what Sundar Pichai said about the platform when the deal came together. Pichai called Intrinsic "the Android of robotics.”

Seven months later and we see the Android playbook unfolding.

  1. Give away the layer everybody needs
  2. Make it genuinely good, genuinely open, genuinely permissive
  3. Let an ecosystem grow on top of it
  4. Keep the parts that make money

Intrinsic is pretty direct about that last part. Solutions built with Intrinsic Core, the company said in its announcement, “automatically work with other Intrinsic offerings, including our enterprise services which includes advanced AI models, Intrinsic Flowstate and industrial grade cloud services. No refactoring or code rewrites needed.”

In other words, while the control framework is free, the digital twin is free and the drivers are free, the advanced AI models are not. Flowstate is not. And the industrial-grade cloud is not.

That’s pretty similar to how Red Hat built a multibillion-dollar company. Essentially, Google is seeding a distribution channel for Gemini-powered physical AI, at a low-level layer where a developer’s architectural decisions get locked in.

Open half, meet closed half

Intrinsic Core is just part of Google’s physical AI push.

Google DeepMind shipped Gemini Robotics 2 on July 30: three models covering whole-body humanoid control, embodied reasoning and on-device execution. Prominent robot makers Apptronik, Boston Dynamics, Agile Robots and Franka were among the launch partners.

Access is fairly tight. The reasoning model, Gemini Robotics ER 2, is available in Google AI Studio and separately in private preview on the Gemini Enterprise Agent Platform, and the vision-language-action model and the on-device variant go to early-access partners by application. The Gemini Robotics SDK requires joining a trusted tester program.

All of this makes sense if you’re Google.

Drivers and kinematics solvers are expensive to build but impossible to defend; programming additional versions of them doesn’t really add value to anyone, least of all Google. Foundation models, however, can be very valuable indeed. So Google has opened the first and gated the second.

The play here might be that every robot running Intrinsic Core becomes a socket a Gemini model can plug into later. But there’s a catch: the two halves don’t visibly touch yet.

DeepMind’s Gemini Robotics 2 announcement makes no mention of Intrinsic or of ROS. The Gemini Robotics SDK doesn’t mention either one. The February announcement said Intrinsic would use Gemini models and work closely with DeepMind, and Intrinsic’s press release today says Core solutions work automatically with its enterprise services “which includes advanced AI models.”

But so far it’s unclear what that interface actually looks like.

Maybe that’s because they’re pointed at different classes of robots. Gemini Robotics 2 is humanoids and dexterous two-armed manipulation, while Intrinsic Core is industrial arms loading CNC machines, with FANUC and Universal Robots support. Is there a meeting point?

I’ve asked Google, and will update this post with the company’s answer.

The market that might benefit most

The machine tending release is a pretty big pointer toward who Intrinsic actually wants.

In the U.S. and Europe alone, Intrinsic says, there are tens of thousands of businesses that fabricate parts, and only 8% have some form of automation. Many of these are small manufacturers that could never afford a six-figure systems integration contract.

This is a genuinely underserved market, and traditional industrial automation has failed it for decades. It’s just been too expensive at lower scale.

That said, prices for robot arms and other forms of automation are coming down, and with key elements like this now available “for free,” this announcement could be a catalyst to opening up the comparatively mom-and-pop segment of automation.

The ROS question, near four years later

There’s history here that robotics veterans will remember.

In December 2022, Intrinsic acquired Open Source Robotics Corporation , the commercial arm of the organization behind ROS, the Robot Operating System that most of the robotics world is built on. ROS, Gazebo, ROSCon and TurtleBot stayed with the nonprofit Open Source Robotics Foundation. Brian Gerkey, who co-created ROS and ran Open Robotics, became Intrinsic’s CTO.

The robotics community was, understandably, nervous about what an Alphabet subsidiary might do with that much influence over open infrastructure.

Gerkey’s answer at the time was essentially: judge us by what we do. A year later, asked by TechCrunch whether ROS was safe under Alphabet, he described the standard Intrinsic would apply to open sourcing its own work: “When we think about what makes sense to be open sourced, we really want to look for the things where we’ve produced something internally that we think is really going to be a game changer.”

Intrinsic Core is kind of the test case.

It’s worth noting that Intrinsic cut 20% of its workforce in January 2023, and there’s a fairly large graveyard of discontinued Google projects. That’s why it’s good news that Intrinsic is open-sourcing some of its software: building your production stack on someone else’s open source is safer than building it on their API.

But open source software for robots is getting to be a bit of a crowded space. NVIDIA has been pushing open models and frameworks into LeRobot with Hugging Face, and Isaac GR00T weights are public. Physical Intelligence open sourced π0 weights and fine-tuning code. Intrinsic’s differentiation is that this stack has been running in actual factories, which not everyone can credibly claim.

It will be interesting to see how – or if – this connects to Google’s other robotics projects from DeepMind for humanoid robots.