By adopting a genetic algorithm for optimization of neural network,the gravity compensation for robot arm is researched.Firstly,through the basic knowledge of mechanics and D-H parameter to set up robot kinematics model,the theoretical computation formula of the gravity item of each joint torque in robot arm is got.Secondly,in the Solid Works simulation software,the simulation value of the gravity item in some certain pose got,thus the correctness of the theoretical computation formula is verified.Finally,the predicted value of the gravity item by genetic algorithm for optimization of neural network is obtained.Experimental results show that the predicted value of gravity items of robot arm learnt with this method is basically conforming with the theoretical value.Consequently,the work load of calculation for the gravity items of robot arm is effectively reduced.Furthermore,this method provide a possible way for the real-time control.