SSTG-Nav: Reusable Metric-Semantic Memory for Object Navigation

A new research paper introduces SSTG-Nav, a reusable metric-semantic memory system designed to improve object navigation for service robots operating in familiar environments over extended periods. Unlike traditional ObjectNav approaches that treat navigation as one-shot exploration, SSTG-Nav leverages a one-time survey to create actionable object goals, consolidating evidence across multiple viewpoints and retaining spatially distinct recovery standoffs. This addresses the challenge of recognizing an object without identifying a reachable stopping point and mitigating the risk of a single map error terminating the task. The system was evaluated on 1,000 HM3D-v2 episodes across 36 scenes, achieving a 99.4% geometric success ceiling with its goal-independent topology. When semantic responses were held fixed, metric grounding improved Success Rate (SR) and SPL from 0.835/0.560 to 0.920/0.603, and source-aware fusion reached 0.926/0.586. Fusion-aware Top-3 recovery further enhanced Success@1/2/3 metrics. The paper is available on arXiv under the identifier 2608.00527.