Answers are grounded in this paper and cite their sources.
Greenhouse controllers record commands but cannot confirm whether actuators execute them. We propose Vision as Witness, a framework that uses vision-language models (VLMs) to infer actuator states from greenhouse video and treat camera observations as visual evidence of actuator behavior. To support this task, we introduce the Naju Actuator Vision-Log Dataset (NAVL), comprising 242 days of controller logs, 149.8 h of video from two cameras, and 1000 manually annotated segments covering irrigation, circulation fans, and a shading curtain. Analysis of the 75 paired days shows that fixed-schedule recording captured only 27.6% of 1557 commands. Zero-shot VLMs perform near chance level (44.5–60.3% balanced accuracy), whereas a spectral motion detector reads the fan at 73.0% and reference photographs in the prompt raise irrigation from 50.0% to 63.0%. With class-balanced LoRA fine-tuning, curtain, fan and irrigation states are read at 75.8–88.1%, 71.4–90.5% and 79.4–95.5% on held-out dates. Importantly, video surpasses a single frame by 52.2, 29.0 and 54.2 points on curtain motion, fan and irrigation in Qwen2.5-VL-7B, at 13 times the prompt tokens. These results suggest that greenhouse cameras can serve as independent witnesses for some actuator states, although they require adaptation for robust verification.
No takes yet. Share an insight, caveat, or question.
Anh et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: