Skill2Real Achieves 78.75% Success on Real-Robot Manipulation via Agentic Skill Learning
Xincheng He · hf · 2026-10-05
Skill2Real is an agentic policy framework that learns executable robot skills through a shared API and transfers them zero-shot from simulation to reality.
- A Proposer-Verifier-Governor loop uses privileged simulation evidence to diagnose outcomes and validate updates while keeping skills grounded in public observations and API semantics.
- A Cerebellum acquires local manipulation skills; a Brain then learns task-level composition with the Cerebellum frozen; both transfer without task-policy fine-tuning.
- GPT-5.6 Sol's frozen checkpoints evaluated by GPT-6 Astra raise LIBERO-Pro Long success from 2.0% to 56.3%; Robosuite training reaches 85.1% (Sol) and 89.4% (Opus 5) mean success across seven tasks; frozen skills achieve 78.75% mean completion on four real-world tasks.
- Removing the Verifier or Governor cuts final Pro Long success by 17.3 and 13.3 points respectively.
More from Embodied
- Asimov shares its journey through the deep robotics stack on Robotics Stack — freelerobot · 2026-10-05
- Six robot arms dancing in sync: peak automation triggered by GPT Image — tsi_org · 2026-10-05
- Zoox robotaxi halts in Vegas traffic despite 3 lidars saying the road was clear — jamesdouma · 2026-10-05
- Show-Harness lets a VLM drive real robots via a physical interface — jiqizhixin · 2026-10-05
- LinkerHand O30 dexterous hand performs card tricks, team to reveal how it was done — DJiafei · 2026-10-05
- Webcam-to-Pen-Plotter Demo at AI Tinkerers Tokyo: Photo to Ink Drawing via AxiDraw — Stefania_druga · 2026-10-05