Skill2Real: Agentic Skill Learning for Zero-Shot Sim-to-Real Robot Manipulation
Skill2Real learns sim-to-real robot manipulation skills via a Proposer-Verifier-Governor loop, achieving 78.75% task completion on real robots without fine-tuning.
Skill2Real is an agentic policy framework that learns executable robot skills through a shared API, using a Proposer-Verifier-Governor loop with privileged simulation evidence to validate updates. A two-level design trains a Cerebellum for local manipulation and a Brain for task composition, transferring both to real hardware without fine-tuning. Using GPT-5.6 Sol on LIBERO-90 with GPT-6 Astra evaluation raises LIBERO-Pro Long success from 2.0% to 56.3%, and Robosuite training reaches 85.1% and 89.4% mean success with Sol and Opus 5 across seven tasks. Removing the Verifier or Governor during training drops final Pro Long success by 17.3 and 13.3 percentage points respectively.
- Proposer-Verifier-Governor loop grounds learned skills in public observations and API semantics
- Frozen LIBERO-90 skills reach 78.75% mean completion on four real-world manipulation tasks
- Ablations show Verifier and Governor contribute 17.3 and 13.3 points on LIBERO-Pro Long
- Organizations
- GPT-5.6 SolGPT-6 AstraOpus 5
From the original article
1 min read at huggingface.co
Transferring robotic skills from simulation to reality requires task knowledge that remains usable across differences in perception, dynamics, and embodiment. We introduce Skill2Real, an agentic policy framework that learns executable skills through a shared application programming interface (API).
By Xincheng He, Siyu Ma, Chang Yu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.