Aiming at the problem of the dynamic model of a cubical robot is a multivariable system and with strong coupling. A balance controller is proposed based on the active diturbance rejection control. Decoupling control of the multivariable system is reallized by introducing virtual control variables, and embedding multi active disturbance rejection controllers(ADRCs) between control variables and outputs. The dynamic coupling and external disturbance are regarded as total disturbances in each channel of the active disturbance rejection controller. After arranging the transient dyanmics of the desired attitudes, we design an extended state observer to estimate and compensate for the total disturbances simultaneously. The parameter-turning of the controller based on experience-based approximation technique and bandwith turning method. A series of experiments are conducted by simulation, and comparative analysis between the ADRC and PID controller is made. Simulation results show that the ADRC not only can control the balance of the cubical robot effectively, but also with superior performance in response rate, control accuracy and strong coupling to the PID controller.