The Vancouver startup says its 'physical AI' software, which controls robots doing contact rich physical work, hit 99.5% success on a 2.5 second cycle plugging wires into connectors on a live automotive line at an unnamed Tier 1 parts supplier.
Sanctuary AI, the Vancouver robotics startup once known for building its own humanoid robot, is making a public bet that the durable asset in industrial robotics is the software brain, not the metal body. The company announced that its "physical AI" foundation model, AI that controls robots performing contact-rich physical work rather than just perception or planning, hit what Sanctuary calls "world-class performance" on a wire-plugging task at a global Tier 1 automotive parts supplier, measured against the customer's own live production benchmarks.
The pivot matters because it repositions Sanctuary in a market where the value chain is being renegotiated. Industrial robotics has long been a hardware business: arms, grippers, controllers, integrators. The newer question is which company owns the AI model layer that makes those arms useful on a real factory floor, and whether that model can move between a startup's own humanoid and the installed base of industrial robots already on customer lines. Sanctuary is now saying yes, and pointing to a benchmark to back it up.
The benchmark is narrow but technically demanding. A robot had to insert a flexible wire into a connector on a part moving along a live conveyor. Flexible wires deform during handling, the connector position shifts as the part travels, and the robot has only a small window to seat the wire before the cycle ends. Sanctuary reports a success rate above 99.5% at a 2.54-second cycle time, numbers the company says it measured against the customer's production benchmarks. The Robot Report, the trade publication that first reported the result, treats the result as a company-validated proof of concept rather than an independent test (Sanctuary AI validates physical AI performance at Tier 1 automotive supplier).
Several limits deserve attention. The task is one task, on one supplier's line, measured by the company building the software. Generalization to other connectors, other parts, other materials, and other line layouts is unproven, and Sanctuary does not claim it. The Tier 1 customer is not identified in the public reporting, which means there is no on-the-record customer voice confirming the cycle time or the production context. The Robot Report's framing, that the numbers are Sanctuary's own against a single customer benchmark, should be the default reader posture until the supplier or an outside researcher confirms independently.
The strategic logic is what makes the announcement worth more than a single benchmark. Olivia Norton, Sanctuary's co-founder and chief technology officer, framed the company as chasing performance and cycle time first, with the assumption that the same AI platform will scale to "the next generation of general-purpose systems." That phrasing matters: Sanctuary is signaling it does not need the world to buy its humanoid hardware for the company to capture value. If a Magna, a ZF, or a Toyota Tsusho can run Sanctuary's model on their own robotics, the addressable market for the software alone is larger than the humanoid market will be for years.
That positioning arrives in a tightening capital cycle for humanoid robotics and a crowded field of competitors making analogous "foundation model for robots" claims. Figure, Apptronik, 1X, Tesla's Optimus group, and Physical Intelligence are all arguing, in different ways, that the model is the moat. The interesting question is not whether Sanctuary's 99.5% figure is real, but whether the company can build a business by licensing the brain to robot owners who have no reason to care which body it runs on. Wire-plugging on a moving conveyor is one credible proof point. The rest of the thesis is still ahead.