High-precision visual monitoring of robot arm position based on deep learning

Abstract

The calibration methods of robotic arms generally have problems such as high cost and low efficiency. Taking the end position measurement link of the robotic arm, which is the most time-consuming and costly among them, as an entry point, a high-precision and real-time visual monitoring method of end position based on deep learning is proposed. Firstly, the influence of the complex error sources of the vision system on the end target position measurement results is analyzed. Secondly, a neural network structure is established based on this, whose input feature is the target position containing the measurement error, and the output label is the accurately computed end joint position. Finally, the training samples and the test samples are generated in the neighborhood space of the specified monitoring point, respectively. Simulation results show that the trained network structure can achieve a prediction accuracy of 0.025 mm on over 99% of the test samples, which is comparable to the laser tracker ranging accuracy. The new method combines the powerful learning capability of neural networks with the complex error sources of vision systems to enable low-cost machine vision methods to achieve high accuracy measurements.