Sanctuary AI published a new technical demo this week showing zero-shot in-hand manipulation on its hydraulic robotic hand. The company says the policy was trained entirely in simulation, then transferred directly to hardware where the hand repeatedly reorients a lettered cube to a target orientation.
That specific task might look narrow, but it sits right at the core of a hard robotics problem: dexterity under contact-rich, high-degree-of-freedom control. In practice, this is the kind of capability ceiling that decides whether humanoid hands can move from “interesting demo” to useful industrial manipulation.
The most important claim here is not just that the hand can do the movement once. It’s the sim-to-real transfer quality. If policies can move from simulation to real hardware with minimal tuning, iteration speed improves dramatically and the path to broader task libraries gets shorter.
This is still not the same as proving deployed, high-uptime production work at scale. But as a technical signal, it’s meaningful: Sanctuary appears to be pushing on the exact bottleneck that many humanoid programs still struggle to cross.
Why this matters
- Better sim-to-real transfer can compound development speed across many manipulation tasks.
- Strong in-hand manipulation capability is a prerequisite for practical general-purpose handling, not just pick-and-place.
- It reinforces a market trend: leading humanoid teams are now competing on dexterity quality, not only locomotion or demo polish.