Answers are grounded in this paper and cite their sources.
Goal-oriented navigation requires an embodied agent to reach a target specified by visual or textual input, integrating perception, spatial reasoning, and decision-making. However, existing methods typically rely on object labels or evaluate relationships between object pairs independently, limiting their ability to capture the overall relational context. We propose Context-Aware Navigation (ContextNav), which models goal verification as a graph-level similarity estimation problem. This enables the agent to determine whether its current observation matches a target goal represented as a scene graph by capturing structure-aware relational dependencies. Unlike previous methods that decompose scene graphs into independent subject-predicate-object triplets, our approach captures dependencies across interconnected relations in the full scene graph. ContextNav includes a Graph-Structured Similarity Estimation (GSSE) module that verifies goal satisfaction by comparing goal and observation graphs generated via VLM-based scene graph generation. It also employs an Online Two-Stage Exploration (OTSE) strategy, which first identifies candidate rooms, computes a minimum-cost inter-room route to the selected candidate under the current room graph, and then performs local goal exploration. Thus, the room-level route is optimal with respect to the current graph and traversal costs, although the complete goal-oriented trajectory is not guaranteed to be globally path-optimal. To evaluate goals specified by explicit object-relation structures, we formulate GraphNav as a structured-goal setting within goal-oriented navigation and introduce ProcTHOR-10K-Context, which extends ProcTHOR-10K with object-context graph annotations. We evaluate ContextNav on four goal modalities: object-, image-, text-, and graph-guided navigation, using HM3D, RoboTHOR, ProcTHOR, and our proposed ProcTHOR-10K-Context, showing consistent gains, particularly on context-rich graph-guided navigation.
No takes yet. Share an insight, caveat, or question.
Phueaksri et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: