HGWM: Hierarchical Graph-Guided World Model for Zero-Shot Object Navigation via Graph Matching

摘要变化

Zero-shot object navigation requires an embodied agent to choosereason about where toa searchtarget may be found while retaining evidence from previous observations. We present HGWM, couplesa anhierarchical LLMgraph-generatedguided world model that combines a category-conditioned room--anchor prior with persistent spatial-semantic memory. A coarse stage assignsuses visuala vision-semanticlanguage scoresmodel to score candidate headings on a 0-10, scale;while a fine stage measures symbolic compatibilityexplicitly betweencompares observed objects, roomand context,relations andwith the goal prior. Phase-dependent fusion combinesintegrates these scoressignals with an LLMa directionlanguage-model recommendation before local action selection.to Onguide HM3Dlocal v0navigation.1, the reported HGWM configuration achieves 62.6\% success rate (SR) and 31.5\% success weighted by path length (SPL), comparedon withHM3D 58v0.1%, improving over WMNav by 4.5 and 310.2%3 forpercentage WMNavpoints, respectively. On MP3D, the corresponding valuesit areachieves 46.8\% SR and 17.9\%, comparedSPL, with 45.4%gains andof 17.2%1. The results indicate higher success point estimates4 and modest path-efficiency0.7 gainspoints. On HM3D v0.2, the full model obtainsreaches 74.4\% SR and 31.9\% SPL.; Additionalremoving componentindividual ablationscomponents reducereduces SR by 0.9--2.3 percentage points and SPL by 0.6--1.5 points. These are author-reported aggregate measurements. The accompanying implementation repairs passresults offlinesupport tests;the episode-leveluse rerunof recordsrelational arepriors notand available.persistent Evaluationgraph ismemory confinedfor totarget-directed simulationsearch.