Illustration brief the model wrote
A robot hand holding a pen mid-air, fingers precisely rotating it. On one side, a blurred human hand making a coarse gesture. In the background, a faint grid and simulated objects fade in. No text, faces, or letters.
DexterityGen: Foundation Controller for Unprecedented Dexterity
Pretrains dexterous motion primitives via sim-to-real RL, distills into foundation controller, uses human teleoperation as coarse prompt, improving hold stability 10-100x and enabling pen/syringe/screwdriver use.
What changed
ModelPreviously, dexterous manipulation required task-specific policies or dense human teleoperation. This paper introduces a single foundation controller that generalizes across tasks and objects, enabling long-duration in-hand manipulation and complex tool use with minimal human input.
So what, in plain English
ModelInstead of programming every finger movement, the robot learns a broad set of hand skills in simulation. A human only needs to give a rough motion hint, and the robot fills in the precise finger actions, like a dancer following a vague direction to improvise the details.
Contribution to industry
ModelThis work contributes a pretrained foundation controller deployable on physical robot hands for assembly, lab, and service tasks requiring tool use. It reduces per-task engineering effort, allowing vendors to quickly enable long-duration in-hand manipulation and robust handling of objects like pens and syringes.
Typed judgments
Jev★★★★☆ manipulation · Physical AI 98%
Real hardware 86% · Open code / data 10% · Foundation-model based 41% · Survey / benchmark-only 2%
Probabilities from TypeSafe Jev over the title, venue and abstract; the filters on the atlas apply code-owned thresholds to these numbers.
Editor's notes
Summary. Pretrains large-scale dexterous motion primitives with sim-to-real reinforcement learning, distills them into a foundation controller, and uses human teleoperation only as a coarse prompt at run time, improving hold stability 10-100x and enabling pen, syringe and screwdriver use.
Industry. Moves tool use and long-duration in-hand manipulation within reach of robot vendors targeting assembly, lab and service work that needs human-grade hands.