Abstract:
The integration of AI-driven methodologies into robotics has become pivotal for advancing automation in high-precision manufacturing, where robotic arms face persistent challenges in controlling absolute positioning accuracy due to nonlinear and random errors. While traditional neural network-based error prediction models offer potential, their performance is often hindered by measurement noise in training samples, leading to slow convergence and suboptimal accuracy. This research introduced a novel Fuzzification-Back Propagation Neural Network (F-BPNN) model, designed to enhance robotic arm positioning accuracy through training data preprocessing by fuzzification and positioning error predicting by neural network. The proposed model leverages fuzzification to preprocess joint rotation angles and directions, transforming raw data into error contributions that mitigate the impact of measurement uncertainties. By integrating a two-dimensional fuzzification process, based on error and error fluctuation, the model optimizes input data for training a BPNN, enabling efficient nonlinear mapping between joint errors and positioning errors of robotic arm’s end. This AI-driven approach not only simplifies samples used by network training but also accelerates training by reducing parameter sensitivity to noise. Experimental validation on one 6-DOF UR5 robotic arm demonstrates significant improvements: training time is reduced by >50 %, mean square error of the model decreases by 2.94 %, and average absolute positioning error of the UR5 robotic arm after compensation is slashed by 59.22 % (from 0.206 mm to 0.084 mm), comparable to existing real-time compensation standards. The model can be easily transplanted into embedded systems to provide a methodology for new design of robotic arm error compensators. By bridging fuzzification with neural network, this work exemplified how hybrid AI techniques can address critical challenges in robotic automation. The results underscore the potential of composite neural network models to predict error in industrial robotics, paving the way for smarter, self-correcting robotic systems in next-generation manufacturing, particularly in high-precision domains like aerospace assembly and microelectronics. Future research will focus on refining fuzzy interval partitioning and expanding the model’s applicability to complex and sophisticated multi-joint collaborative robots, further aligning with the vision of AI-driven automation as a cornerstone of Industry 4.0.
Biography:
Jiang Liu is a Ph. D candidate in the School of Aerospace at Harbin Institute of Technology (HIT), Heilongjiang Province, China. He earned his Master Degree of Engineering from HIT in 2022 and Bachelor Degree of Engineering from Wuhan University of Technology, Hubei Province, China, in 2020. During his graduate studies, he worked on AI-driven solutions for real-time positioning error compensation in industrial robotic arms and contributed to the development of high-precision 2D motion stages. Currently, his research interests include positioning error prediction in industrial robotic arms, fuzzification-neural network modeling, and advanced packaging technologies for integrated circuits, with an emphasis on wafer hybrid bonding techniques.
Jiang Liu
