Abstract: Industrial robots are affected by multiple sources of uncertainty during their whole life cycle, which leads to deviations between the actual and desired end-effector positions, affecting the quality of the products being processed. For this reason, a study on error compensation in industrial robots has been carried out, in which a full-domain refined error field modelling method and error compensation strategy based on the principle of non-kinematic calibration are proposed and experimentally verified. It aims to provide an effective tool for improving the accuracy performance of domestic industrial robots. The radial basis function network is trained with the nominal position of end-effector as input node and the corresponding position error as the output node. Then, the cross-validation and particle swarm optimization algorithms are used to improve the training efficiency and model accuracy. Thus, a position error field of industrial robot in full domain is accurately constructed with few samples. According to the well-developed error field model, the error for any position point of industrial robot can be predicted. After that, error compensation is realized by editing the controller setpoints to increase the pre-bias. The error compensation experiments are conducted for three kinds of domestic industrial robots with rated loads of 3 kg, 12 kg and 50 kg, respectively. The proposed error field model has higher accuracy compared with the error model based on error similarity principle using the same sample points. After compensation, the absolute position error ranges of the experimental industrial robots are reduced by 44.14%, 77.48% and 80.65%, the maximal absolute position error are reduced by 42.55%, 76.07% and 82.24%, verifying the engineering applicability of the proposed method.

Key words: industrial robot, error compensation, error field model, radial basis function neural network, accuracy durability