Answers are grounded in this paper and cite their sources.
Object goal navigation involves navigating to an object of a specified target category in a previously unseen environment. During exploration, the accumulated observations provide spatial and semantic context for subsequent navigation decisions. We propose a graph-based world model with a graph value memory, which organizes observed locations and unvisited navigation candidates into a structured representation. Conditioned on this memory, the world model predicts candidate-specific future room semantics and target relevance to select navigation directions. On a scene-disjoint Matterport3D (MP3D) validation split, the world model head achieved mean top-1 and top-3 accuracies of 61.53% and 81.57%, respectively. In an evaluation of 400 MP3D validation episodes, the proposed method achieved a success rate (SR) of 35.0% and a success weighted by path length (SPL) of 12.10%, compared with 32.0% and 6.79% for the baseline. With the ground-truth object list held constant, removing either the world model or the graph context given to the path-planning vision-language model lowered SR and SPL (32.0% and 10.15%; 32.2% and 7.68%, respectively).
No takes yet. Share an insight, caveat, or question.
Kyeong-Ho et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: