• A error estimate model is established and verified by testing to predict the pose error of the robot. • Combines the advantages of deep nerual network and error similarity, the statistical feature of robot error similarity can improve the robustness and prediction accuracy of the error estimate model. • Cross contrast training can extract the contrast information and enlarged the sample size. • The proposed scheme doesn't depend on the kinematic parameters of the robot and needs neither retrofitting with the robot's control system and massive measurements. • The proposed scheme can effectively reduce the maximum positional errors and orientation errors of the robot. Abstract With the advantages of the high degree of freedom and large action space, industrial robots are gradually widely used in high-end large-scale equipment automatic assembly fields such as aerospace. It is very important to improve the absolute pose accuracy of the robot. This paper presents a deep learning scheme for the compensation of absolute pose errors for industrial robots. A mapping model of the robot's theoretical pose coordinates and actual pose errors is established based on the deep belief networks and error similarity, the pose error prediction is realized. Combined with the offline feedforward compensation method, the scheme has been verified on a 6-DOF serial industrial robot KUKA KR500-3 of robot automatic assembly system. The proposed scheme combines the feature expression of deep learning with statistical interpretability of error similarity, does not depend on the specific kinematic model of the robot, and without retrofitting with high-end encoder. After compensation, that the maximum absolute positional error and orientation error of the robot end-effector was reduced by 83.99% from 1.524 mm to 0.244 mm and reduced by 54.88% from 0.082 deg to 0.037 deg respectively, which can meet the accuracy requirement of the robotic automatic assembly system.