A multi-agent collaboration framework for scene understanding and work order generation in smart farms

Le résumé fourni par la source

Manual inspection is insufficient for large-scale smart farming, which requires near-real-time monitoring and closed-loop task management. This study proposes a multi-agent framework for scene understanding and automatic work order generation in cattle farms. To handle complex environments, large illumination changes, and low-frequency abnormal events, we design a state-flow-driven architecture with five agents: visual perception, semantic analysis, rule evaluation, temporal confirmation, and work order generation. The framework integrates multimodal large language models, using a fine-tuned InternVL3.5-8B model for visual perception and scene understanding and DeepSeek-V3.2 for semantic reasoning, rule-based evaluation, and task generation. A structured state object enables stepwise decision-making across layers. We further introduce a consecutive-frame confirmation mechanism to impose temporal consistency and reduce false triggers caused by single-frame errors. A structured work order model maps detected events to executable tasks, forming a closed loop from perception to action. We evaluate the method on a real farm dataset covering indoor and outdoor scenes, multiple time periods, camera views, and task categories, with frame-level manual annotations. Results show that the proposed framework is stable and robust in complex farm conditions and can reliably detect abnormal events and generate work orders automatically.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.