The paper discusses a multi-agent system for robot-soccer. The system consists of a supervisory controller, and controllers for attack, defense and goalie robots. Real time vector field based path planning, Petri-net theory and Q-learning technique are used in the design. Robot-soccer system has a dynamic environment. A soccer robot has to take an appropriate decision based on environment situation. The supervisor assigns the role of each robot. Experimental study verifies applicability and efficiency of the developed controller scheme.