Skill2Real: Agentic Skill Learning for Zero-Shot Sim-to-Real Robot Manipulation
Skill2Real reports frozen simulation-trained robot skills transferring to real manipulation without real-world fine-tuning.
The paper describes an agentic policy framework built around a shared robot API and a Proposer-Verifier-Governor loop. Its hierarchy learns local skills first, then task-level composition, with both carried over unchanged to the real robot. In the reported results, frozen LIBERO-90 skills reached 78.75% mean completion across four real-world manipulation tasks. HF Daily Papers' note
The paper describes an agentic policy framework built around a shared robot API and a Proposer-Verifier-Governor loop. Its hierarchy learns local skills first, then task-level composition, with both carried over unchanged to the real robot. In the reported results, frozen LIBERO-90 skills reached 78.75% mean completion across four real-world manipulation tasks. HF Daily Papers' note
score 5