Institute for Physical AI @ BMI · a demonstration
The rule from the last lesson tells you if a grip will hold — but something has to look at a shape and choose where to grab, out of every possible grip. Here that something is a neural network — a real one, scoring candidate grasps around the object on the board and picking the best, then lifting it. Watch it choose. Then guess how big the brain that does this actually is.
Provenance. The policy is a 4→16→16→1 multilayer perceptron, behaviour-cloned from the exact force-closure quality: an approach angle and the object's shape in, a predicted Q out. At run time it scores a fan of candidate grasps and the gripper takes the best — a nano version of a learned grasp model (DexNet). Released open. The demonstration — a policy you can watch judge grasps on-device, and count the cost of — is ours.