Artificial Neural Network Guided Compensation of Nonlinear Payload and Wear Effects for Industrial Robots

In the paper, we have developed a novel hybrid approach to calibrate industrial robots and improve their positioning accuracy. As robots are increasingly used for high-precision manufacturing, sub-millimeter accuracy is required. However, both geometric errors from inaccurate kinematic parameters and complex non-geometric errors cause overall tool-center-point positioning errors on the order of millimeters. Our key insight was to combine the traditional kinematic model based on Denavit-Hartenberg (DH) parameters with a neural network.

The DH-based kinematic model relates joint angles to tool position via transformations defined by link parameters. However, it cannot account for non-geometric errors from payloads, wear, and other factors. The neural network compensates for these complex errors by predicting the error between planned and actual tool positions given input joint angles and relevant data like the mass of the payload. We trained our networks using measurements collected from a Universal Robots UR5 operating under different payloads.

We tested our hybrid approach through experiments and simulations. For payload compensation experiments, optimizing the DH parameters followed by neural network training reduced errors by up to 59% compared to the robot’s default calibration. Our joint wear simulations showed the hybrid approach could maintain residual positioning errors below 0.25mm over hundreds of simulated wear cycles.

By using kinematic optimization to first reduce geometric errors, our neural network could focus on non-geometric errors. This achieved substantially better accuracy than approaches relying only on one model. Our robot-agnostic hybrid calibration provides a promising way to enhance industrial robot accuracy.