Sanctuary AI's June robot-brain strategy targets factory arms it does not build

Why it matters

Sanctuary AI can pursue industrial contracts without waiting for a Phoenix fleet. Success would turn years of embodiment research into a control layer for robots factories already own.

On June 17, Sanctuary AI co-founder Olivia Norton laid out an expanded strategy: run Sanctuary AI's Physical AI on industrial arms that factories can buy today, instead of tying every installation to Phoenix, the humanoid Sanctuary AI spent years building.

In an August 29 interview with Forbes, CEO Daniel Friedmann supplied the important qualification. The automotive work behind Sanctuary AI's headline performance numbers remains a proof of concept. Sanctuary AI is packaging its control software, data infrastructure and end effectors for industrial customers, but its disclosed test is not a production deployment.

That distinction turns what looks like a robot product update into a revealing founder decision. Norton previously led Kindred AI's Vancouver artificial-general-intelligence group before co-founding Sanctuary AI in 2018. She now oversees engineering and operations, carrying forward a research program that began with an ambitious premise: human-like intelligence would require a human-like body.

Sanctuary AI is keeping that premise while relaxing the customer's hardware requirement. Factories can start with existing industrial robot arms, according to Sanctuary AI's product materials. Phoenix can wait for humanoid hardware, manufacturing and economics to catch up.

The numbers need two clocks

Sanctuary AI announced in June that its system had achieved a success rate above 99.5% while inserting flexible wired plugs into moving targets for an unnamed Tier 1 automotive supplier. Sanctuary AI reported a 2.54-second cycle time measured against the customer's live production benchmark.

Friedmann told Forbes that the test ran for 40 minutes and included 313 insertion trials, with a 99.5% success rate. Taken across the full test window, that works out to an average of about 7.7 seconds per trial. Sanctuary AI's 2.54-second number therefore describes the measured task cycle, a narrower metric than total elapsed test time.

Buyers evaluating the system will care about both. Cycle time determines whether a robotic process can keep pace with a production line. Elapsed throughput captures stoppages, resets and other time surrounding each insertion. Sanctuary AI has reported the benchmark and test-window figures, while the automotive customer has not been named.

The trial was also small enough to keep the claim in perspective. A reported 99.5% success rate across 313 attempts amounts to roughly one or two failed insertions, depending on the underlying count and rounding. It establishes repeatability during a 40-minute test, rather than reliability across shifts, factories or changing components.

Sanctuary AI has previously described teleoperation as a way to collect demonstrations for training autonomous robot behavior, and its current product page presents the wire task as live policy performance without edited attempts. The measurements and validation remain company-generated.

A founder thesis gets a cheaper body

Sanctuary AI grew out of research conducted at Kindred, the robotics developer co-founded by Suzanne Gildert and Geordie Rose. Gildert has written that Kindred's remaining work on human-like robots and artificial general intelligence was spun out into Sanctuary AI in 2018. Norton, Rose, Gildert and University of Toronto economist Ajay Agrawal founded Sanctuary AI around that effort.

The technical lineage runs through quantum computing and warehouse robotics. Rose co-founded D-Wave and later Kindred. Gildert worked as an experimental physicist at D-Wave before building teleoperated robots at Kindred. Norton brought computer engineering and biomechanics experience, including work at ETH Zurich, before leading Kindred's Vancouver AGI group.

Gildert's original embodiment thesis pushed Sanctuary AI toward increasingly human-like machines. Her account of Sanctuary AI's early years describes an initial attempt to build robots with human-like faces, followed by a decision to focus engineering resources on functional hands and bodies when facial mechanics proved distracting.

The industrial-arm strategy is the latest narrowing of that scope. Sanctuary AI can preserve years of work on perception, touch, manipulation and policy training while removing locomotion, balance and full humanoid manufacturing from the first customer installation. The product page currently foregrounds industrial arms with low-degree-of-freedom grippers and custom end effectors. Sanctuary AI lists its hydraulic robotic hands as a future configuration for this offering.

The factory task gives Sanctuary AI a concrete result to take into sales conversations without asking a plant operator to become an early humanoid adopter.

Friedmann inherits the commercialization job

Norton is pursuing that shift alongside Daniel Friedmann, Sanctuary AI's CEO and chairman. Friedmann previously spent nearly two decades leading space-technology developer MDA and later ran Carbon Engineering. His appointment places an executive with experience commercializing capital-intensive hardware beside the co-founder who has carried Sanctuary AI's engineering program since 2018.

Friedmann told Forbes that Sanctuary AI aims to deploy the industrial system within weeks, depending on task complexity. He declined to provide a typical return on investment and confirmed the automotive example had not entered production. The near-term commercial test is straightforward: whether a manufacturer will pay Sanctuary AI to reproduce its proof-of-concept performance under continuous factory conditions.

Sanctuary AI has financial room to pursue more than one embodiment, although its current revenue and pricing have not been published. In July 2024, Sanctuary AI said it had attracted more than C$140M in total investment after financing from BDC Capital's Thrive Venture Fund and InBC Investment Corp. Sanctuary AI named Accenture, Bell, Export Development Canada, Evok Innovations, Magna International, SE Health, Verizon Ventures and Workday Ventures among its other backers. Zeon Ventures later invested and formed a materials partnership, without announcing financial terms.

The robot brain market is already splitting

Sanctuary AI's expanded scope puts it against two different strategies. Figure is integrating its Helix control models with its own humanoid hardware; its Helix 02 system controls locomotion, manipulation and balance across the full Figure robot. Physical Intelligence is developing models intended to transfer across bodies; its pi0 research used data from eight robot configurations.

Sanctuary AI is choosing a middle course shaped by its founders' history. It continues to build proprietary hardware, sensors and hands, while offering the intelligence layer on robot bodies supplied by established industrial manufacturers. That gives Sanctuary AI access to factories before Phoenix becomes a standard piece of equipment and gives customers a way to test its software without replacing an entire automation stack.

The strategy will be judged on production contracts, sustained uptime and deployment costs. For now, Sanctuary AI has demonstrated a narrow task at a production-relevant speed, disclosed enough detail to examine the result and stopped short of calling the proof of concept a factory deployment. That is a more credible starting point than another humanoid video looking for a job.