In this paper, we propose a Neural Network learning architecture for the reflex-control of complex systems. The multi-layer-network is trained by a new learning algorithm which does not need a desired output, but directly minimizes a criterion which spezify the control objective. We propose two learning concepts: off-line learning and on-line learning (similar to adaptive control). We have tested our method for a simplified dynamic problem: the dynamic stability of a simulated planar biped. To obtain dynamic equilibrium during the single-support-phase, the robot trunk is controlled such that a zero-moment-point (ZMP) criterion is satisfied. We show that although no gait-specific knowledge is used during off-line learning, dynamic walking for different step lengths and heights could be realized and controled. The capability of the proposed learning algorithm to adapt the trunk control on-line to the actually performed leg trajectories allows further improvements and adaptation to real environment conditions or perturbations.