Answers are grounded in this paper and cite their sources.
Vision-Language Models (VLMs) offer strong semantic reasoning for robotic scene understanding, but their high computational cost makes continuous inference impractical for real-time navigation. We propose an efficient vision-based triggering framework that gates expensive VLM inference using object-level change evidence, activating VLM reasoning only when task-relevant changes suggest potentially abnormal situations. To improve reliability in mobile robotic environments, we introduce a target-aware change detection strategy that evaluates changes only for predefined task-relevant objects, substantially suppressing false alarms caused by viewpoint variations and irrelevant scene dynamics. Our method is built on a single YOLOX backbone that unifies view recognition for retrieving viewpoint-aligned reference images and a lightweight cross-attention-based change head for object-centric change prediction, avoiding additional heavy modules. Experiments on real and synthetic benchmarks demonstrate a strong accuracy–efficiency trade-off; for example, our model runs at 10 FPS with 6 M parameters and achieves an AUROC of 0.838 on a QRB5165 robotics platform, outperforming prior change detection baselines that use substantially larger backbones.
No takes yet. Share an insight, caveat, or question.
Baek et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: