Abstract This paper presents a novel framework for modeling role and task allocation in cooperative wheeled soccer robot systems by leveraging latent knowledge extracted from past collaborative interactions. Inspired by recent advances in heterogeneous multi-robot collaboration, the proposed method encodes a soccer team as a set of Multidimensional Relational Structures (MDRSs), capturing both temporal and spatial relations among robot roles, actions, and stimuli. A structured dataset, termed the Soccer Robot Collaboration Treebank (SRCT), is introduced to represent play-by-play histories of robot behaviors, parsed through a formal grammar to support structured learning. Probabilistic modeling and Non-Negative Tensor Decomposition (NTD) are applied to the resulting tensors, enabling robust inference and latent knowledge estimation even in scenarios with sparse data or communication loss. Simulated experiments using a team of wheeled soccer robots in the Webots environment demonstrate the system’s ability to dynamically reassign roles, reason over incomplete histories, and predict collaborative behaviors such as passing, defending, or role-switching. The results show that the proposed framework enhances both strategic flexibility and robustness, providing a foundation for real-time decision-making in robotic soccer under uncertainty.