A new dataset called COSI-Lab provides multimodal, multi-sensor recordings of a real interdisciplinary scientific workshop held at an international conference. The dataset captures 32 academics as they engage in naturally occurring social interactions, offering what the authors describe as ecologically valid data for studying social behavior.

The recordings were made in a weakly scripted setting, meaning participants were given only minimal guidance. This allows the dataset to reflect realistic conference dynamics while still enabling controlled analysis. The goal is to support research on modeling social intention from multiple perspectives and modalities.

The work is described in a paper posted on arXiv, where the dataset is introduced as a resource for studying social intention in complex, real-world environments.