It is necessary to identify and compensate the non-linear friction to improve the performances of the electro-hydraulic system of robotic excavators. The mathematical model of the electro-hydraulic system is established first, and the non-linear friction are described by the improved Stribeck model. The actual friction force is calculated by measuring the pressures of the two chambers of the hydraulic cylinder at different constant velocities. The fitness, crossover and mutation probabilities of the genetic algorithm are improved to overcome the shortcomings of stagnation and premature convergence. Four unknown parameters in the friction model are identified and verified by the improved genetic algorithm. According to the displacement signal from the sensor, a velocity observer is used to estimate the velocity signal in real time. Based on the principle of structural invariance, a dynamic friction feedforward compensation method is designed. On this basis, the experiments of friction compensation (FC) under three working conditions of sinusoidal trajectory, low and high velocities tracking were carried out for the arm electro-hydraulic system. The results show that the proposed Stribeck model and friction compensation control method effectively eliminates the low velocity crawling and amplitude flattening phenomena, and improves the control performances of the system.