Sanctuary AI built a humanoid. Its best industrial worker so far is a pair of hands.

Automating the un-automatable?

Walk through a modern automotive assembly plant and almost everything you see is robotic. Welding, painting, stamping, torquing, material handling ... all were solved decades ago, in some cases twice over.

But not the wiring.

Wire harnesses are the single most stubbornly manual part of building a car. And there are a lot of them. The reason is pretty obvious: a wire is what roboticists call a deformable linear object. It flops. It changes shape when you touch it. It has no fixed geometry to grab, no repeatable pose to program against, and the whole hundred-billion-dollar edifice of industrial robotics is built on the assumption that the thing you're picking up will be in the same place, in the same shape, every single time. 

But can innovations from humanoid robots ... errr ... save the day?

Wire harness assembly is still done by hand because deformability defeats the methods that work on rigid parts. But not anymore.

That's the task Sanctuary AI just claimed.

Earlier this summer the Vancouver humanoid robotics company announced a 99.5%-plus success rate at a 2.54-second cycle time inserting flexible wired plugs into a moving target on a live conveyor at a global Tier 1 automotive supplier. Benchmarked, it says, against that customer's existing line performance.

But, surprise, surprise: there was no humanoid robot involved. Just Sanctuary's Physical AI, driving arms and hands, embedded in somebody else's production line.

In other words, a company that spent years building a body is discovering that its most employable product right now is the brain, along with a decent pair of hands. Which, in fairness, is where Sanctuary always said the hard part lived. Former CEO Geordie Rose told me that "half of the complexity of the robot is in the hands." As it turns out, you can sell that half separately.

The timing is also better than it looks. Plugging cables into things is quietly becoming one of the most wanted robotic skills on earth ... and not only in car plants.

"Meta is testing robots that can plug in cables, reset servers, and handle other tasks inside its data centers, according to several current and former workers familiar with the projects," a recent Wired story says.

That's the same fundamental problem. Floppy cable, precise target, no tolerance for error, and an industry that would very much like to stop hiring humans to do it at 3 a.m.

So Sanctuary's strategy now reads like this: rather than wait for humanoid hardware to reach mass commercialization, deploy its existing Physical AI and gripper/hand solutions inside existing commercial industrial processes and robots. That's hardware-agnostic, available today, and — the company argues — the foundation for the industrial humanoids that come later. It's also something that will live on even after great humanoid robots ship, because you don't always need (or want) a full body.

But great manipulation is always going to be useful somewhere.

You can still find Phoenix, Sanctuary's full humanoid, on the company website. Generation 8 offers a strong torso, arms and head on a wheeled base, with significant improvements in manufacturability and sensor data. But the top-line promise on the home page now reads "physical AI for industrial automation," and the images and video with the highest prominence show arms, hands and grippers attached to or embedded within production lines.

So: is this a pivot? A side quest? Or an early appetizer on the path to a full suite of robotic and automation tools for a wide variety of tasks?

I mean, not every humanoid robot maker on the planet is going to succeed, right?

I put the question to Daniel Friedmann, who took over as Sanctuary's CEO in late June. He moved up from board chair, replaced interim CEO James Wells, and arrived with a résumé built on scaling hard-engineering businesses: longtime chief executive of space robotics firm MDA, and later the clean-energy company Carbon Engineering.

He had some answers.

John Koetsier: How was the 99.5% success rate measured?

Daniel Friedmann: We tallied successful insertions against unsuccessful ones to calculate the overall success rate.

John Koetsier: Over how many production cycles was that performance achieved?

Daniel Friedmann: The test was for 40 minutes and included 313 plug insertion trials.

John Koetsier: How does the 2.54-second cycle time compare to human workers and existing automation?

Daniel Friedmann: This task has not been possible to automate with traditional methods. The 2.54-second cycle time was benchmarked against the customer's existing line performance.

John Koetsier: Is this a strategic shift away from humanoids?

Daniel Friedmann: This represents an expanded approach to Physical AI and solving critical labor challenges. Our hardware-agnostic capabilities expedite industrial adoption of Physical AI, while building the foundation that will support the next generation of intelligent robotic systems, including industrial humanoids.

John Koetsier: Why deploy your AI on industrial robots instead of the Phoenix humanoid?

Daniel Friedmann: Rather than waiting for humanoid hardware to reach mass commercialization, we're deploying Physical AI on existing commercial platforms today. This lets us deliver production-ready performance to customers now, while building toward the next generation of intelligent robotic systems, including industrial humanoids.

John Koetsier: What did building humanoids teach you that enabled this deployment?

Daniel Friedmann: Global industrial leaders are facing unprecedented labor shortages and increasing operational costs. When it comes to solving those challenges, function matters more than form.

John Koetsier: Do you consider Sanctuary primarily an AI company or a robotics company?

Daniel Friedmann: Sanctuary AI is a full-stack company: we're developing both Physical AI and robotic hardware. We need access to the best hardware to train our Physical AI with the highest quality data, and we need highly capable Physical AI to utilize that hardware to its fullest extent. The relationship between hardware and Physical AI is symbiotic.

John Koetsier: Is this deployment already in production?

Daniel Friedmann: No, this is at the proof-of-concept stage.

John Koetsier: How broadly can this system be deployed across manufacturing tasks?

Daniel Friedmann: Our Physical AI is designed to handle the contact-rich, dexterity-intensive tasks that have historically been out of reach for traditional automation. For companies in manufacturing, logistics, and other labor-constrained industries, it's deployable on existing robotic hardware today, giving customers production-ready performance and a clear path to the next generation of intelligent industrial systems.

John Koetsier: How long does it take a customer to deploy the technology?

Daniel Friedmann: Our goal is to deploy within a few weeks. Currently, the timelines vary based on the complexity of the task.

John Koetsier: What is the typical ROI for customers?

Daniel Friedmann: We are focused on maintaining customer throughput at a lower cost, but it's too early to share specifics beyond that.

John Koetsier: Does the system work with existing robot hardware?

Daniel Friedmann: Yes. Our Physical AI can run on currently available industrial robots using both off-the-shelf end effectors and custom ones designed by our team. We also intend to support the next generations of intelligent robotic systems as they come to market, including industrial humanoids.

John Koetsier: Thank you for your time.

What the numbers do and don't say

A few things worth thinking about.

Forty minutes is a demo, not a shift. At 99.5% across 313 trials, we're talking about one or two failures an hour. So that needs some work.

And, the solution needs to work for 8-hour shifts or longer.

The bet underneath

Co-founder and CTO Olivia Norton put the thesis plainly in the original press release:

"Physical AI adoption is gated by AI that meets both performance and cycle time requirements. By focusing on a performance-first approach to Physical AI models, we're delivering value to customers today."

Gated by the AI. Not by the body.

If that's right, a big chunk of the humanoid race is being run on the wrong axis, or at least the wrong emphasis. Getting the job done, after all, is more important than HOW you get it done. And the winners will be whoever's models hit production reliability first — on whatever hardware happens to be available.

The company has raised more than $140 million, cut roughly 30 staff in late 2024 after co-founder departures, took an investment this spring from the venture arm of Japan's Zeon Corporation, and has now cycled through three CEOs in short order. Handing the job to someone who spent decades scaling hard-engineering businesses — MDA built NASA's Canadarm — is a reasonably clear signal about which phase the board thinks the company is in.

Function matters more than form, Friedmann says. It will be interesting to see what functions Sanctuary ends up offering.