Hand–eye matrix estimation via motion similarity: a novel multi-source fusion method for eye-to-hand calibration in robotics
Xiaoyao Wang, Chang Liu, Yicong Chen, Fuzhou Du · Measurement Science and Technology · 2025
Abstract Eye-to-hand calibration is an essential technique for achieving precise alignment between vision sensors and robot control systems, and it is widely recognized for its significance in the effective implementation of vision-guided robots in advanced manufacturing. However, traditional eye-to-hand calibration often faces challenges in attaining high precision in hand–eye matrix estimation within industrial applications. These challenges arise from absolute positioning errors in robots, complications related to parameter identification, and the need for expert knowledge. To address these challenges, this paper introduces a novel multi-source information fusion framework, along with a new hand–eye matrix estimation model based on this framework, referred to as differentiable generalized regression neural network (diff-GRNN). The proposed method eliminates the need for parameter identification and expert knowledge, automatically estimating the unknown hand–eye matrix by leveraging motion similarity. Experimental results indicate that the proposed method is promising for eye-to-hand calibration in robots under varying error conditions, achieving greater accuracy compared to other prominent calibration techniques. The translational error in the calibration results is reduced from 2.5071 mm using traditional methods to 0.3087 mm with the proposed approach.