This paper describes a control structure for planar humanoid runner on uneven terrain utilizing output zeroing. In this paper, the terrain whose inclination changes at random within a certain range is called uneven terrain. First, the change of dynamics by running on uneven terrain is considered, and the change of dynamics by running on uneven terrain is negated by adding the impulse input. The running on uneven terrain whose inclination changes within plusmn7 [deg] is achieved by the simulation. Second, important parameters of the output function is modified online utilizing Q-learning to acquire the running motion which can dynamically cope with the inclination of terrain. As a result, the running on uneven terrain whose inclination changes gradually within plusmn5[deg] is achieved.