The authors have been using the ZMP (Zero Moment Point) as a criterion to distinguish the stability of walking for a biped walking robot that has a trunk. In this paper, the authors propose a learning control algorithm of the compensative trunk motion that makes the actual ZMP get closer to the desired ZMP. The convergency of the algorithm is confirmed by computer simulation. By the learning experiments with the biped robot, the reappearance of the measured ZMP is shown and the convergency of the algorithm is confirmed. The change of the convergence rate with the change of the weight coefficient multiplied to the errors between the measured ZMP and the disired ZMP is confirmed by the simulation and the experiments. And also the reasons are discussed.