New GADAR policy enables zero-shot generalization in classical planning

Researchers have developed GADAR, a novel policy for classical planning that can generalize across different domains without requiring search. Unlike previous methods that tie policies to specific domains, GADAR takes the domain description as input and constructs grounded actions in unseen domains. This is achieved through a three-layer approach: a lifted domain layer for structural signatures, a binding layer for action applicability, and a chain layer to track schema dependencies. In evaluations, a single GADAR policy trained on seven benchmark domains achieved 58.3% zero-shot generalization, performing nearly as well as dedicated single-domain models. AI

IMPACT This research could lead to more adaptable AI planning systems that require less domain-specific training.

RANK_REASON The cluster contains a research paper detailing a new method for classical planning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• CatalyzeX Code Finder for Papers
• DagsHub
• Hugging Face
• Rajesh Mangannavar
• Visitall

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →