A Novel Self-recoverable Hand-eye Calibration Technique for Vision-guided Robot

Abstract

Long-term operability is essential to reap the benefits of cost-effective industrial automation involving vision-guided robots (VGR). However, it is essential that the accuracy of the robot during the period of operation is uncompromised. The hand-eye calibration accuracy of a VGR plays a huge role in the accuracy of the robot during operation as it enables the proper perception of the environment in which a VGR operates. Hand-eye calibration of VGRs is typically done offline. However, the calibrated parameters may change during the operation, which requires recalibration to maintain operational accuracy, therefore taking the robot offline resulting in operational downtime and lost revenue. As such, it becomes imperative to integrate a system that enables the robot to maintain its accuracy by adapting to changes in its environment. To tackle this challenge, different hand-eye calibration techniques for vision-guided robots were reviewed to assess and evaluate their strengths and weaknesses in an operational environment, as well as the practicality of different types of calibration targets. Furthermore, experimental validation of the different factors that affect the accuracy of hand-eye calibration was carried out using six commonly used algorithms in the industry. To minimise the problems common with these algorithms, a self-recoverable hand-eye calibration scheme based on a hybrid filter is proposed. This algorithm runs simultaneously with the robot’s operation while ensuring the robot recovers from any changes in the calibration accuracy. The hybrid filter is based on a combination of Kalman and particle filters with optimisations on particle transition and genetic algorithm-based resampling as a proposed modification.To reduce the dimension of the particle filter for improved performance, the particle filter, which estimates the rotation parameter,is coupled with a Kalman filter with iterated state update to estimate the translation parameter. While the proposed particle transition scheme handles the problem of sample impoverishment common with the standard particle filter, the genetic algorithm resampling method handles the problem of degeneracy without inducing sample impoverishment. A gradient descent estimator, which has a much simpler implementation than the Kalman filter was also explored for the translation parameter estimate. Experimentation with the algorithm shows that it is able to actively detect and recover from errors due to changes in the calibration parameter. Comprehensive experimental results with a UR5e robot arm show that even in the offline scenario, the proposed method outperforms the offline-specific calibration method, highlighting the suitability of the developed method for online and offline calibration

Date of Award	May 2024
Original language	English
Awarding Institution	Coventry University
Supervisor	Babul KSM Kader Ibrahim (Supervisor), Ranveer Matharu (Supervisor), Mathias Foo (Supervisor) & Hafiz Ahmed (Supervisor)