Sanctuary AI brings learned manipulation to existing industrial robots

Sanctuary AI develops learned control for robots that handle variable physical work. Its current commercial emphasis is Physical AI on established industrial robot arms, with task-specific end effectors. That is an important change of emphasis for a company associated with the Phoenix humanoid: today’s deployment route should be evaluated separately from its hydraulic hands, mobile manipulation and future humanoid roadmap.

01 / ProductThe present offer puts learned policies on industrial hardware

Sanctuary’s Physical AI platform describes data collection, policy development and deployment across workflows. It identifies FANUC and Universal Robots among the industrial hardware it integrates with and supports custom or off-the-shelf grippers. The buyer is evaluating intelligence applied to a work cell, not necessarily purchasing a new humanoid body.

The embodiment roadmap makes the availability boundary unusually clear. Industrial arms with custom end effectors are labelled as deploying now. Proprietary hydraulic hands are in development, mobile platforms are described as piloting soon, and humanoids are a future destination. Those stages should remain separate even when they appear in the same company presentation.

A June 2026 announcement explains the strategy using a wire-plugging proof of concept with an unnamed Tier 1 automotive supplier. Sanctuary reports performance against that customer’s production benchmarks. This is evidence of a particular vendor-reported demonstration and strategic direction, not an independently established guarantee for every factory task.

The product idea is to learn the variable parts of manipulation that are awkward to express as a fixed motion programme. A wire bends, a part arrives slightly displaced and contact provides information that a camera alone may miss. Learned behaviour still needs a defined robot, tool, work area and completion criterion. The AI sits within those physical constraints.

02 / AudienceManufacturers with variable manipulation have a concrete starting point

Sanctuary’s insertion material targets work involving wires, connectors, pins and harnesses. Its kitting page describes sorting and placing varied parts into complete kits. These are useful reader situations because the challenge is specific: handle variation while maintaining the required orientation, quantity or connection, at a pace useful to the next production step.

An automotive supplier may investigate a connector operation that is difficult to fixture consistently. An electronics manufacturer may consider kit preparation with frequent part changes. In both cases, the strongest candidate task has a clearly defined acceptable result and enough recurring volume to justify collecting data and integrating a work cell.

An organisation that already owns industrial robots should not assume existing equipment is automatically reusable. Ask Sanctuary to identify the supported controller, arm, sensing and end-effector combination for the task. Reusing a familiar robot brand can reduce some uncertainty, but the complete installed configuration must still be checked.

A reader seeking a generally available humanoid employee should interpret the current roadmap cautiously. The company’s work on dexterous hands remains relevant to future systems, but it is not the same commercial proposition as the arms-and-grippers offer. A near-term factory project should stand on the capabilities and support available for the proposed installation now.

03 / WorkflowA proposed connector trial separates learning from acceptance

Consider a proposed pilot for inserting a flexible connector into a moving assembly. Sequenced has not tested Sanctuary’s system. Begin with the physical definition of success: the correct part, seated to the specified position, with no damage and with whatever downstream verification the production process requires. A robot completing a motion is not sufficient evidence that a connection is acceptable.

The insertion page describes learning from human demonstrations and force-aware control. For the proposed pilot, the process engineer should select demonstrations that include expected variation in presentation and material behaviour. Data from only the easiest assembly condition would create a misleading picture of the operating problem.

Keep some production-representative examples separate from policy development. Evaluate the trained behaviour on those examples, including awkward but allowable starting positions. Record completed acceptable assemblies, elapsed time, retries and operator interventions. Analyse failures by type: poor approach, incomplete seating, damaged component or an upstream presentation problem may need different corrections.

Test the surrounding sequence as well. The cell must receive the next assembly, know which variant is present and communicate completion to the line. If a connector cannot be seated, the agreed response may be a retry, rejection or operator assistance. That response should be part of the trial, rather than improvised after the robot encounters its first difficult item.

Finally, repeat the accepted task after a planned changeover or restart. This checks whether the process is usable by production staff and whether task configuration is sufficiently explicit. The desired result is a repeatable operation with a known exception path, not a single impressive success rate detached from ordinary factory conditions.

04 / PricingA scoped industrial engagement, with two ways to develop tasks

The current commercial inquiry asks about hardware interest, anticipated fleet size and automation priorities. The reviewed pages do not publish a universal robot price, software subscription or per-task charge. Treat the commercial model as a scoped engagement whose hardware, policy development and service terms need to be quoted.

Both kitting and insertion describe two development routes: the customer’s team can develop tasks through the platform, or Sanctuary’s solutions team can develop the initial task with the customer. The phrase self-serve describes how a task may be developed; the pages do not establish anonymous signup, a free tier or immediate access without a commercial relationship.

This distinction changes the economics. A capable automation team may want tools and training so it can extend an accepted task. Another factory may prefer the supplier to own more of the initial development. Ask which work is included, who maintains policies after a product change and how another line or site is priced.

Compare the total cost to accepted work produced over the intended operating period. Keep robot hardware, end-effectors, integration, data collection and support visible in the proposal. The public evidence does not tell a buyer whether these are bundled or separately billed. A clearly scoped quote is more useful than a speculative hourly price for a future humanoid.

05 / DistinctionsThe roadmap separates intelligence from the eventual body

The hydraulic hand programme describes tactile sensing and a high-degree-of-freedom hand intended for industrial manipulation. Its page also says pilot programmes are forthcoming. That makes it a development direction worth following, while preventing a near-term buyer from assuming the hand is a standard shipping component of an industrial-arm engagement.

The present strategy is significant because many useful tasks do not require legs or a human-shaped torso. A fixed arm can work at a station where the objects already arrive. If the difficult part is adapting a grasp or insertion, improving that behaviour can deliver value before a more ambitious body becomes suitable for production.

The Figure blueprint provides context for a humanoid-centred path into industrial work. Sanctuary’s current offer asks a different question: can learned policies make existing industrial embodiments useful on a specific task? Neither approach eliminates the need to establish the operating envelope, but they allocate the hardware-development problem differently.

The Skild AI blueprint is relevant when examining the relationship between robotic intelligence and different bodies. Sanctuary’s buyer still needs a concrete hardware configuration and a task-development engagement. A broad ambition for reusable intelligence should not replace evidence about the exact process the factory is considering.

06 / QuestionsA proof of concept is evidence with a defined boundary

Sanctuary’s June announcement reports a 99.5%+ success rate and 2.54-second cycle for its particular wire-plugging task. Those are vendor-reported results from a named task context, not Sequenced measurements. Before applying them to another line, ask about the evaluation period, number of trials, part variation and what counted as success or intervention.

Failures that look rare in a percentage can matter at production scale. What happens to the failed item, and how long does recovery take? A process may tolerate a rejected component but not an unnoticed incomplete assembly. Evaluate detection and recovery together with average success; the right threshold depends on the actual manufacturing requirement.

The roadmap also leaves a practical question for buyers interested in future embodiments: what can be contracted and supported now? Request a written separation between the current cell, any pilot hardware and future possibilities. A roadmap can guide architecture choices without becoming a dependency for the initial project’s business case.

Data ownership and iteration deserve an explicit discussion. The site describes customer collaboration on task data and policies, but the reviewed public pages do not settle every right to reuse demonstrations or move trained work to another installation. Establish those terms before relying on a planned multi-site expansion.

07 / DecisionEvaluate the useful industrial task available today

Sanctuary AI is relevant to physical AI because it connects learning from demonstration with difficult industrial manipulation. Its current strategy gives buyers a near-term route on established robotic hardware while keeping more dexterous and mobile embodiments on a separate roadmap. That separation makes the company easier to evaluate on work that exists now.

A good first engagement begins with a constrained task, representative variation and an agreed accepted outcome. Decide whether the internal team or Sanctuary will lead task development, then assess the complete process at its required pace. Future hands and humanoids can remain strategic interests without being treated as present-day entitlements.