Seeing Through the Days: A Multiday Office Cubicle Dataset for Associative Embodied VQA

Item association is a core component of scene understanding in real-world environments. People often refer to objects using names, numbers, spatial descriptors, or contextual cues, leading to inherently ambiguous and overlapping associations that are hard for large language models to understand, especially across different time points. In long-term real-world deployments, robots must also resolve these complex and evolving associations to effectively interact with their surroundings. As scenes grow in visual complexity and temporal duration, current vision-language models (VLMs) struggle to maintain consistency, resulting in degraded performance on key tasks such as object detection and classification, localization, state estimation, and counting.
However, these real-world challenges—characterized by clutter, redundancy, and temporal variation—are not well represented in existing benchmark datasets, which are often static, synthetic, or lack significant visual feature redundancy. To address these limitations, we introduce Cubicle4D, a new benchmark dataset that evaluates item association in dynamic 4D office scenes, where objects move across both space and time. Our preliminary analysis reveals a linear degradation in model performance as the number of cubicles increases, underscoring the need for new approaches to address data association in large-scale, dynamic environments.