The trick to a useful robot hand is putting the physics in twice

Robotics is quietly rewriting where the human knowledge goes. Instead of asking one neural network to learn physics from scratch — picking, rolling, twisting — researchers hand the network a head start: the parts of physics that don't change, baked into the hardware and the training objective at once. The skill that has stayed out of reach, rolling an object inside the palm, is one place the approach is paying off.

The new arXiv paper states the recipe plainly: a global grasp-quality prior — a decades-old scoring rule from classical mechanics that judges how well fingers distribute force — combined with a local contact-geometry prior built into the fingertip's curve. The two priors work in tandem; one steers the learning, the other steers the contact.

The repeatable move: a robot skill gets durable when human insight lives in two places at once — the metal you can hold, and the score the optimizer is trying to beat. Pure end-to-end learning — a single network, sensors to motors, no human-built structure — makes the network rediscover gravity on its own. In the paper's framing, co-design hands it a shortcut on the parts of physics that stay the same.

The honest line: one paper, three objects, four palm orientations, author-reported numbers. Independent reproduction will sort method from test bed. The winners in robotics over the next five years will be the teams that learn where to put the shortcut — in the hand, in the score, or in both.