Sanctuary AI is making a specific bet on hand dexterity
Their hand design stands out for a hydraulic approach that differs from the electromechanical and tendon-driven systems more commonly used elsewhere.
Sanctuary AI's latest demo showed its robotic hand reorienting a lettered cube to a target orientation ten consecutive times without dropping it. The company said the policy was trained in simulation and transferred zero-shot to the real hand. The demo also highlighted fingertip-only manipulation rather than using the palm to stabilize the object, which makes the task more demanding.
Another angle here is the hardware choice underneath it; Sanctuary is betting on hydraulics in the hand, which stands apart from the electromechanical and tendon-driven approaches more commonly used elsewhere. The hand has 21 degrees of freedom, including active degrees of freedom for finger abduction and in-hand manipulation.
They tie that approach to miniaturized hydraulic valves and unusually high power density. In their December hand update, Sanctuary said the design gives them an order-of-magnitude power density advantage over cable and electromechanical systems, along with gains in speed, strength, controllability, cycle life, impact resistance, and heat management. The valve actuators had passed 2 billion test cycles without leakage or degradation.
Their recent hand work has not been limited to hardware. Sanctuary says they are using NVIDIA Isaac Lab to train dexterous manipulation policies in simulation and then transfer them onto the real hand. They also say simulation lets them train thousands of hands in parallel and explore hand kinematics beyond what analogous teleoperation can access.
They have been extending the sensing side as well. In February, Sanctuary said they integrated tactile sensors into Phoenix and linked them to blind picking, slippage detection, and force control when vision is occluded or arrives too late. That fits the rest of their hand program, which keeps returning to contact-rich manipulation rather than treating dexterity as a narrow benchmark problem.
The hardware and simulation work reflects sustained IP investment. Morgan Stanley ranked Sanctuary third in published humanoid-related patent filings over the previous five years. The ranking covers the full humanoid stack, but the hand program shows where much of that development effort has gone.