Cooperative multi-agent reinforcement learning aims to train decentralized agents to accomplish joint tasks by maximizing a global reward. While existing value decomposition methods under the centralized training and decentralized execution paradigm have achieved notable success, they often overlook the latent role structures inherent in multi-agent systems. In real-world scenarios, agents may exhibit functional heterogeneity or behavioral diversity, even when sharing identical observation and action spaces. To address this limitation, we propose Role-Aware Dynamic Grouping (RADG), a novel framework that learns contrastive role representations from agents’ trajectory information and performs adaptive grouping based on these learned roles. The extracted role embeddings capture meaningful behavioral patterns that guide flexible and dynamic group formation. Within each group, agents coordinate more effectively through shared policy information and group-aware value decomposition. RADG enables structured cooperation, improves exploration efficiency, and enhances generalization across diverse tasks. Importantly, it operates without relying on manual supervision or domain-specific priors, making it well-suited for dynamic and complex environments. Experimental results on standard cooperative MARL benchmarks demonstrate that RADG consistently outperforms existing baselines in coordination performance, training stability, and adaptability to varying team structures.