Problem
Research questions and friction points this paper is trying to address.
sim-to-real transfer
zero-shot manipulation
robot skill learning
embodiment gap
Skill2Real: Agentic Skill Learning for Zero-Shot Sim-to-Real Robot Manipulation
🤖 AI Summary
This study addresses the challenge of transferring robotic skills from simulation to reality, where perceptual, dynamic, and embodiment discrepancies hinder deployment. To overcome this, the authors propose a cerebellum-cortex hierarchical architecture coupled with a shared API policy framework that first acquires local manipulation primitives before composing complex tasks. Furthermore, they introduce a Propose-Verify-Govern (PVG) closed-loop mechanism that leverages simulated evidence to iteratively diagnose and refine policies, enabling foundation model-driven zero-shot cross-domain transfer without real-world fine-tuning. The proposed approach yields substantial improvements on the LIBERO-Pro Long benchmark, increasing the success rate from 2.0% to 56.3%, while achieving an average task completion rate of 78.75% on physical robots. These results effectively validate the feasibility of hierarchical skill transfer for sim-to-real robotic applications.
📝 Abstract
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). A Proposer-Verifier-Governor (PVG) loop uses privileged simulation evidence to diagnose outcomes and validate updates, while keeping learned skills grounded in public observations and API semantics. The Cerebellum first acquires local manipulation skills; the Brain then learns task-level composition with the Cerebellum frozen. Both memories transfer to the real robot without task-policy fine-tuning or skill-memory updates. As GPT-5.6 Sol learns skills on LIBERO-90, evaluating each frozen checkpoint with GPT-6 Astra raises LIBERO-Pro Long success from 2.0% to 56.3%, without training on Pro Long. Independent Robosuite training reaches 85.1% and 89.4% mean success with Sol and Opus 5 across seven tasks, respectively. Frozen Sol-trained LIBERO-90 skills achieve 78.75% mean completion across four real-world manipulation tasks with Astra. Removing the Verifier or Governor during LIBERO-90 training lowers final Pro Long success by 17.3 and 13.3 percentage points, respectively. These results support learning and transferring a hierarchy of executable skills through a common robot interface.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Agentic Skill Learning
Zero-Shot Sim-to-Real
Proposer-Verifier-Governor Loop
Hierarchical Policy Composition
Shared Robot API
🔎 Similar Papers
No similar papers found.
