In the area of autonomous multi-robot cooperation, much emphasis has been placed on how to coordinate individual robot behaviors in order to achieve an optimal solution to task completion as a team. This paper presents an approach to cooperative multi-robot reinforcement learning based on a hybrid state space representation of the environment to achieve both task learning and heterogeneous role emergence in a unified framework. The methodology also involves learning space reduction through a neural perception module and a progressive rescheduling algorithm that interleaves online execution and relearning to adapt to environmental uncertainties and enhance performance. The approach aims to reduce combinatorial complexity inherent in role-task optimization, and achieves a satisfying solution to complex team-based tasks, rather than a globally optimal solution. Empirical evaluation of the proposed framework is conducted through simulation of a foraging task.