SFC-VLM: Single-Frame Compression of Temporal Information for Robotics Vision-Language Model

Abstract

Vision-Language Models (VLMs) are effective at static image reasoning, but existing VLM-based safety systems remain biased toward static posture recognition (e.g., lying or sitting) rather than dynamic event understanding. In robotics, this distinction is critical because recognizing a state alone is insufficient to determine whether an event has actually occurred. Assistive robots require continuous monitoring and prompt responses to safety-critical transitions, which demands process-level reasoning. However, single-frame inference lacks temporal context, while video-based multi-frame pipelines are often computationally heavy for low-latency, on-device deployment. To address this gap, we propose SFC-VLM (Single-Frame Compression for Vision-Language Models), a framework that transforms temporal reasoning into spatial reasoning by compressing consecutive frames into a single image for event-level understanding. SFC-VLM employs two compression methods: Grid Compression (time represented as spatial order) and Motion Trail Compression (time represented through progressive fade/blur), along with compression-aware prompting that guides the model to leverage spatial reasoning for interpreting temporal sequences. We evaluate SFC-VLM on the Le2i Fall Detection Dataset using frame-level classification at 10 FPS. Grid Compression with a $2 \times 3$ grid layout achieves 58.06 F1 score, representing a 40% improvement over the no-compression baseline (41.49), while maintaining 97.00% specificity. Latency analysis shows that Grid Compression achieves an average event response time of 0.465 s and an average detection latency of 0.158 s, suggesting its potential as a low-latency temporal compression approach for robotic perception with only one VLM inference per decision.