Behav2Need: Ethology-driven animal need understanding via multimodal sensing on smart farms

Authors: WANG, RUIQING - University Of Illinois; ZHANG, JIALE - University Of Michigan; LEE, SUNGMIN - University Of Michigan; Miles, Jeremy; Rohrer, Gary; DONG, YIWEN - University Of Illinois

Submitted to: Proceedings
Publication Type: Proceedings
Publication Acceptance Date: 4/10/2026
Publication Date: 6/25/2026
Citation: Wang, R., Zhang, J., Lee, S., Miles, J.R., Rohrer, G.A., Dong, Y. 2026. Behav2Need: Ethology-driven animal need understanding via multimodal sensing on smart farms. In Proceedings: BuildSys '26 Proceedings of the 13th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation. Banff, Alberta, Canada, June 22-25, 2026. p. 13-23.

Interpretive Summary:

Technical Abstract: Long-term behavioral monitoring is important for smart built environments to understand and adapt to occupants’ needs. However, prior work largely focuses on behavior recognition (i.e., what the behavior is), with limited attention to why and how behaviors occur, which is essential for inferring occupants’ needs and providing targeted support. This limitation is particularly critical in animal-centric environments, where occupants cannot explicitly communicate their true needs. To address this challenge,we develop Behav2Need, an ethology-driven framework leveraging Dynamic Causal Bayesian Networks for animal need understanding through multimodal behavior sensing in smart farm environments. The key insight behind ethology is that behavior emerges from temporally structured interactions between internal needs and external environmental stimuli. Modeling these inherent causal relationships enables the inference of latent needs, more accurate behavior prediction, and need-aware intervention. There are two key challenges. First, the misalignment between high-level behavioral mechanisms and real-world noisy, asynchronous, and indirect multimodal sensor data. Second, the causal relationships between latent needs and behaviors are context-dependent and evolve over time. We address these challenges by extracting variables and estimating empirical priors from sensor data collected from animal units based on animal science domain knowledge, and by allowing causal relationships and probabilistic causal effects to be iteratively refined using real-world data. We evaluate our approach through a real-world smart-farm deployment with 82 pigs across 8 farrowing crates over 1,152 hours, achieving 73% behavior prediction accuracy and generating multiple interpretable causal graphs conditioned on time × behavior combinations to improve understanding of animal behavior and its underlying contextual and ethological mechanisms.