GAVEL: Graph World Model Lifts Qwen3-8B Long-Horizon Robot Planning from 41% to 92%
Ruiyang Wang · hf · 2026-09-21
GAVEL is a framework for verifying and repairing long-horizon LLM robot planning built around an explicit graph world model capturing object relations, action preconditions/effects, and probabilistic beliefs over unobserved object locations. It predicts action consequences before execution, detects violations, repairs directly fixable errors, and reserves LLM replanning for semantic reasoning. On BEHAVIOR-1K (100 single long-horizon tasks, 500 multi-task instructions), it raises Qwen3-8B single-task success from 41.2% to 91.8% and multi-task success from 19.9% to 92.6%, while distributional belief reasoning cuts travel distance by 5.4%.
More from Embodied
- Walden Robotics' Sage model quadruples robot uptime on Toyota factory lines — adnothing · 2026-11-12
- Robotics debate: matching training to test distribution is not the best approach, says researcher — EpisodeYang · 2026-11-12
- Deployment is where physical AI grows up — and its most important data source, per Tedrake's CoRL talk — m_wulfmeier · 2026-11-12
- Runway teases triple bet: real-time avatars, interactive worlds, robot action models — c_valenzuelab · 2026-11-12
- WICL: robots learn manipulation in-context, 83.4% success on all 50 RoboTwin tasks — _akhaliq · 2026-11-12
- After CoRL, robotics feels far from solved: Gemini 4 Carbon evaluated — DJiafei · 2026-11-12