This paper is concerned with the generation of a balancing trajectory for improving the walking performance. The balancing motion has been determined by solving a second -order differential equation. However, this method caused some difficulties in linearizing and approximating the equation and had restrictions on using various balancing trajectories. The proposed difficulties in linearizing and approximating the equation and had restrictions on using various balancing trajectories. The proposed method i this paper is based on the genetic algorithm for minimizing the motins of balancing joints, whose trajectories are generated by the fifth-order polynomial interpolation after planning leg trajectories. The real walking experiments are made on the biped robot IWR-III, developed by our Automatic Control Laboratory. The system has 8 degrees of freedom and the structure of three pitches in each leg, and one roll and one prismatic joint in the balancing joints. The experimental result shows the validity and applicability of the new proposed algorithm.