Count the actuators and the story changes — and Sanctuary put most of its actuator budget in the hands. US12172317B2, granted December 2024, is "Artificial intelligence-actuated robot," and it couples dexterous hand hardware tightly to learned control.
The B25J filing tells a different story than the keynote. Classified under B25J 9/1664 (vision-guided manipulation), B25J 15/10 (gripping hands) and B25J 13/085 (tactile sensing for control), the patent is about a hand that is both mechanically dexterous and controlled by AI that integrates touch. The 'AI-actuated' framing fuses the hardware and the policy that drives it.
Sanctuary's bet is the inverse of the locomotion-first orthodoxy. Where Boston Dynamics' lineage obsesses over staying upright, Sanctuary argues that walking is increasingly solved and the real moat is fine manipulation — hands that can do the fiddly bimanual work that makes a humanoid economically useful in a warehouse or a back room.
The honest constraint is the one every dexterous-hand program hits: a human hand has roughly two dozen degrees of freedom packed into a power-dense, durable, sensor-rich package, and replicating even a fraction of that in a robot hand that survives industrial duty is brutally hard. Dexterity is a battery problem before it's an AI problem, and a hand densely actuated enough to be useful is a thermal and power problem too.
Pairing this with Sanctuary's companion 2024 grant on teach-through initialization (US12157226B2) shows the full thesis: dexterous hands plus efficient teaching. The hand does the hard physical work; the teaching system makes new tasks cheap to add.
For readers auditing the humanoid race, Sanctuary marks one pole of a real strategic split. Locomotion-first or manipulation-first is a genuine fork, and the patents tell you which fork a company took. Sanctuary filed for the hands. Whether the hands are where the value lands is the wager the whole field is running.