This article presents a method for cooperative robot actions through multi-stage construction of action intelligence. The intelligent composite motion control is a learning methodology for intelligent robots that gradually realizes complex actions from fundamental motions. It is not easy, however, to obtain the action intelligence of a complex cooperation at once, hence the intelligence is constructed through multi-stage action intelligence optimization. For optimization multi-stage genetic algorithm, MGA, is used. The MGA solves a large-scale optimization problem with complicated constraints as multi-stage combinatorial optimization problems with simple constraints, which are solved by GA to generate their suboptimal solution sets. The empirical knowledge obtained is effectively utilized in similar situations and moreover to achieve more complex actions. The method is successfully applied to optimal realization of cooperative robot soccer actions. And it is shown that the resultant action intelligence has large applicability.