AI
Skill2Real: Agentic Skill Learning for Zero-Shot Sim-to-Real Robot Manipulation
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 01/10/2026, 20:00:00 UTC
Briefing created: 05/10/2026, 02:37:37 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 05/10/2026, 02:18:24 UTC.
Briefings
Skill2Real is an agentic framework enabling zero-shot simulation-to-real robot manipulation by learning executable skills via a shared API. The Cerebellum acquires local manipulation skills, while the Brain learns task-level composition with the Cerebellum frozen. This approach transfers learned skills to real robots without fine-tuning, achieving 56.3% success on LIBERO-90 with GPT-6 Astra and 78.75% mean completion across four real-world tasks. Key findings include significant performance gains when Verifier and Governor components are removed, demonstrating the robustness of the learned hierarchy.
Key points
- Skill2Real employs a Proposer-Verifier-Governor (PVG) loop using privileged simulation evidence to diagnose outcomes and validate updates while maintaining skills grounded in public observations.
- Frozen Sol-trained LIBERO-90 skills achieve 78.75% mean completion across four real-world manipulation tasks with Astra, demonstrating zero-shot transfer without task-policy fine-tuning.
Context and limitations
Results are limited to specific simulation environments (LIBERO-90, Robosuite) and evaluation agents (GPT-5.6 Sol, GPT-6 Astra). Success rates are reported as mean values across multiple checkpoints and tasks, with specific success rates varying by task and agent configuration.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 05/10/2026, 02:37:37 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeibu3n7vfvdjqxurbsdaax2yha5t4yvqgsrrodfe2bxnldcxkzbuva