Sanctuary AI ports its humanoid stack onto industrial arms factories already own

Selling a policy layer onto arms a plant already owns changes what has to hold at the seam. Sanctuary AI keeps the perception, touch, manipulation and policy training; locomotion, balance and full humanoid manufacturing come out of scope ^[14]. The end effectors stay in the package ^[4], which is the honest call: a flexible wired plug going into a moving target is a hand problem before it is a brain problem.

Two clocks matter. Forty minutes is 2,400 seconds, and 313 trials spread across 2,400 seconds is about 7.7 seconds of wall clock each ^[1]. The published cycle is 2.54 seconds ^[6], and 313 of those is 795 seconds, which leaves roughly 1,605 seconds, about 27 minutes of the window, sitting outside the measured cycles ^[2]. Sanctuary AI's own framing agrees that 2.54 describes the task cycle rather than elapsed time ^[13]. The remainder is part presentation, resets, recovery and whatever else happened between insertions, and the source material does not break it down.

For the cycle figure to set line pace, the part has to arrive at the tolerance the policy trained on, the moving target has to keep moving the way it moved that morning, and the dead time has to be absorbed by fixturing rather than by an engineer standing at the cell.

The failure arithmetic is thin in the same way. One insertion is worth 0.319 percentage points of a 313-trial rate ^[3], so the distance between 99.5% and 99.0% is a single plug, and the reported rate corresponds to one or two misses ^[9]. Extend the cycle instead of the window: 2.54 seconds across an eight-hour shift is about 11,338 insertions ^[4], and half a percent of that is roughly 57 failures per shift ^[5]. Whether 57 is tolerable depends on what a bad plug costs downstream, and that figure belongs to the customer, not the vendor. What the trial establishes is repeatability, which is what the source claims for it ^[10].

One more thing to hold in view. Sanctuary AI has previously described teleoperation as a way to collect demonstrations for training autonomous behaviour ^[11], and the current product page presents the wire task as live policy performance with no edited attempts ^[12]. Both statements can be true at once, and the only thing that settles it for a buyer is standing at the cell with a stopwatch.

My read is that this lowers Sanctuary AI's exposure to its own hardware schedule. Olivia Norton, who led Kindred's Vancouver AGI group before co-founding the company in 2018 and now runs engineering and operations ^[15], has put the position as: factories can start with arms they can buy today ^[2], while Phoenix waits for humanoid hardware, manufacturing and economics to catch up ^[3]. The founding premise that human-like intelligence needs a human-like body survives intact ^[16]; the customer just stops paying for the body. I would fund a pilot on this. I would also write the acceptance test against elapsed throughput over a full shift and treat 2.54 seconds as the best case it currently is.