Learning dexterity

OpenAI just showed off something that sounds simple but is brutally hard in robotics: a hand that can hold an object and just... move it around. Rotate a block, reposition it in the palm, adjust grip without dropping anything. Humans do this without thinking. Robots have historically been terrible at it, because hands have dozens of degrees of freedom and physical contact is messy and unpredictable.

The hand in question is a Shadow Dexterous Hand, a five-fingered robotic hand that's been kicking around research labs for years as a benchmark for manipulation. OpenAI's twist wasn't better hardware. It was training the control policy almost entirely in simulation, using reinforcement learning, then transferring that policy onto the physical hand with surprisingly little friction.

That sim-to-real jump is normally where these projects fall apart. Simulated physics never quite match real-world friction, weight, or motor response, so a policy that looks flawless in software often flails once it meets an actual object. OpenAI's answer was to randomize the simulation aggressively — varying friction, mass, and even visual appearance thousands of times over — so the neural network learns a policy robust enough to handle whatever quirks the real hand throws at it, rather than one perfectly tuned to a single idealized simulation.

The result is a hand that can perform in-hand manipulation tasks that trip up most robotics systems, using nothing but camera input and the same kind of trial-and-error learning that's powered game-playing AI. No hand-coded rules about finger placement, no explicit physics models programmed by engineers. Just millions of simulated attempts distilled into a policy that works on physical hardware.

It's a small step toward something bigger: robots that can generalize their physical skills the way large language models generalize text. Dexterity has long been the bottleneck between AI that can reason and AI that can actually do something with its hands in the physical world.