Anatomy of Grasp.
The same idea, a third machine — a hand, taken apart three ways: the fingers that hold, the rule that decides if a grip will slip, and the brain that chooses where to grab. This is the regime the drone and the humanoid never touch: contact. Where they stayed up by balancing forces in open space, this one has to seize the world and not let go. The rule that decides if a grip holds is a scalar certificate — force closure, positive when caught, negative when it will slip — the grasp-space cousin of the energy certificate, in exactly the contact regime the certification frontier still calls open. Every number on the board is computed by the Institute's open grasp core, so the page cannot lie to you — and it completes the trio: fly, balance, grasp.
Guess before you look.
Each opens with a question almost everyone gets wrong, an object you can turn over in 3D, and the one move that makes it obvious. Open them in order, or jump to the layer you want.
How hard does it squeeze?
“How hard must it squeeze to hold a 1 kg cup without dropping it?”
Take the hand apart: two fingers, friction cones, contacts. The surprising number is the squeeze — and it is tiny. Grip force is weight over twice the friction, about the object's own weight, and pressing harder buys nothing. The grip holds by geometry, not strength.
Open the demonstration →Why does a grip slip?
“Two grips, the same squeeze. What decides whether it holds or spins free?”
The one geometric rule you have felt every time a jar lid fought you: force closure. A grasp holds only if its contacts can push back on every twist — a quality Q that is positive when caught, negative when not. The board shows an aligned pinch that holds and an oblique one that spins out, at the same squeeze.
Open the demonstration →What decides where to grip?
“A neural network chooses where to grab. How big is its brain?”
Something has to look at a shape and pick a grip out of all of them. Here it is a network — 369 numbers — scoring candidate grasps around the object live and taking the best, a nano version of a learned grasp model. It learned to predict force closure, then chose a grip that holds on shape after shape.
Open the demonstration →One format, the whole of embodiment.
Three machines now run on the same engine — the same withhold-guess-prove beats, the same editable source, the same rule that nothing is animated by hand. A drone that stays up by pushing air down, a humanoid that stays up by not falling over, and a hand that holds by geometry. Fly, balance, grasp: the three ways a body meets the world, in one format.
Guess first
A real question, refused an answer, until you commit. This one asks how hard a robot must squeeze — and the honest answer, barely at all, is the surprise that opens the door.
It cannot lie
Every claim is checked live by the open grasp core: the grip force set by friction, the force-closure quality that flips a grasp from held to slipping, the 369-parameter policy whose chosen grip holds on 24 of 24 shapes. Verified, not asserted.
Hardware to brain
One hand carries the whole stack: the contact mechanics of the fingers, the force-closure rule that decides if a grip holds, and the learned policy that chooses where to grab — the three layers of an embodied system, a third time.
Start with the body →Take the course: Embodied AI Design →← Anatomy of Balance← Anatomy of Flight