Abstract: Concentrated on the danger of robot turnover caused by sudden external shocks during normal walking,a balance control method based on orbital energy model is presented,which introduces a regression model based on the robot's acceleration and other status information to predict the condition of being turnover,and uses both attitude compensation control and"taking a step"strategies to achieve balance recovery.Meanwhile,an offline learning method is used to model the relation between body state and optimized foothold to optimize theoretic foothold.In the experiment of AIBO robots,by this method,the frequency of turnover is reduced by more than half,locomotion efficiency is improved effectively,and thus the method's validity and feasibility are proved.