Human intergroup coordination in a hierarchical multi-agent sensorimotor task arises from concurrent co-optimization.

Division of labor and specialization are common principles observed across all levels of biological organisms and societies, including humans that often rely on specialized roles to achieve a shared goal in complex coordination tasks. Understanding these principles in a quantitative fashion remains a challenge. In this study, we explore a novel experimental paradigm where two specialized groups of human players-a sensor group and an actor group-collaborate to accomplish a shared sensorimotor task of steering a cursor into a target. With all decision-makers initially unaware of their contribution and in the absence of verbal communication, the study explores how the group dynamics evolve over time, evaluating performance in terms of learning speed, group coherence and intergroup coordination. To gain quantitative insights, we simulate different computational models, including Bayesian learning and bounded rationality models, to describe human participants' behavior. We also relate our findings to perceptual control theory, which emphasizes hierarchical control systems in which information flows bidirectionally between levels. Our results show that both human participants and model-based simulations (Bayesian and bounded rational agents) successfully complete the task. Over time, mutual information between actors and sensors increases, and cooperative behavior emerges within the groups. Interestingly, model-free hierarchical reinforcement learning fails to account for the observed data, being overwhelmed by task variability. In contrast, model-based approaches can be shown to generalize to larger groups and more complex network structures in evolutionary simulations. Our findings highlight the importance of internal models and concurrent co-optimization in facilitating adaptive coordination, offering insights into distributed information processing mechanisms.