Skill2Real: Agentic Skill Learning for Zero-Shot Sim-to-Real Robot Manipulation

Skill2Real. Agentic Skill Learning for Zero-Shot Sim-to-Real Robot Manipulation. By Xincheng He and seven others, from the University of California, Los Angeles, Shanghai Jiao Tong University, and two other institutions. Published on the second of October. No code repository is linked to it.

Moving a robot skill from simulation to the real world means carrying task knowledge across three kinds of difference: in perception, in dynamics, and in the body doing the work. Whatever is learned in simulation has to stay usable once all three have changed underneath it.

Skill2Real is an agentic policy framework whose skills are executable, and learned through a shared application programming interface, an API. The interface is the common ground. A skill written against it in simulation is written against the same thing the real robot exposes.

Learning runs through a loop of three roles: a Proposer, a Verifier, and a Governor, together called PVG. The loop uses privileged evidence from the simulator to diagnose outcomes and to validate updates, while the skills it learns stay grounded in public observations and in the semantics of the API. The privileged information informs the learning without leaking into the skills.

Skills are learned in two tiers. A component called the Cerebellum acquires local manipulation skills first. Then, with the Cerebellum frozen, the Brain learns how to compose those skills at the level of whole tasks. Both memories transfer to the real robot as they are, with no fine-tuning of the task policy and no updates to skill memory.

With GPT-5.6 Sol learning skills on LIBERO-90, each frozen checkpoint is evaluated using GPT-6 Astra. Success on LIBERO-Pro Long rises from two percent to fifty-six point three percent, with no training on Pro Long itself. Separately, training on Robosuite reaches eighty-five point one percent mean success across seven tasks with Sol, and eighty-nine point four percent with Opus 5.

On the real robot, frozen LIBERO-90 skills trained with Sol reach a mean completion of seventy-eight point seven five percent across four manipulation tasks, run with Astra. Ablations show what the loop's checks contribute. Removing the Verifier during LIBERO-90 training lowers final Pro Long success by seventeen point three percentage points. Removing the Governor lowers it by thirteen point three.

The authors take this as support for learning a hierarchy of executable skills and transferring it through a common robot interface. Two details hold the picture together. The skills cross from simulation to the real robot frozen, with nothing retrained on the far side. And the checking roles in the loop are not decoration: take out either the Verifier or the Governor, and the skills that come out of training are measurably weaker.