in this thesis, a structured hierarchical methodology based on petri nets is used to introduce a task model for a soccer goalkeeper robot. in real or robot soccer, goalkeeper is an important element which has a key role and challenging features in the game. goalkeeper aims at defending goal from scoring goals by opponent team, actually to prevent the goal from the opponent player’s attacks. this purpose although is simple but has very important effects in soccer playing. this thesis develops and implements a complete and effective tasks for a goalkeeper soccer robot. all the models are marked ordinary petri nets (mopn) or generalized stochastic petri nets (gspn) which is an extension of petri nets. petri net places represent predicates, actions and tasks and transitions represent events. the petri net based task plan can be analyzed for both qualitative (logical) and quantitative (performance) properties. to increase the performance of goalkeeper, it has been considered some decision points in tasks models so as to select the choice which is better when encountering this points. by applying a learning algorithm to these decision points, goalkeeper will learn to do one of the alternatives, by repeatedly facing with different situations, it converges to do the one which is optimized way by probability 1. results of the qualitative analysis of proposed model and applying the proposed learning algorithm to improve performance of goalkeeper for one base task plan, are presented.