ORCID

Abstract

The adoption of low volume high mix manufacturing lines in global production sites has lead to a great need for flexible automation solutions. Flexible intelligent robots as an example are increasingly deployed in new industrial environments, such as the intra-logistics, in order to increase the automation level of the mentioned manufacturing lines. The lack of structure and the high dynamic of such industrial environments lead to potential performance deterioration related to unexpected events that are not recognized by the robot. Detecting such changes in the robot environment is believed in this thesis to be possible with anomaly detection techniques which allows the robot to react to them either autonomously or with user interaction. In other words, this work examines the impact of integrating anomaly detection in productive intelligent robots on the stability and quality of the automated process. A modular framework, called the post gripping perception framework, has been developed in this research to define a methodical integration of image anomaly detection in intelligent manipulation robots, whose software is based on image object detection for pick and place tasks. Adding anomaly detection modules for the planning, monitoring and verifying of the post gripping steps has been identified to be efficient in enhancing the stability and quality of robotic processes. The framework also defines the nature of anomaly detection approaches needed in the different modules, namely structural anomaly detection in the planning and verifying modules for quality enhancement and logical anomaly detection in the monitoring and verifying modules for stability increase. The challenge of collecting rare anomaly data in productive robotic processes has been considered in this thesis by introducing an approach to transform state-of-the-art semi-supervised anomaly detection into unsupervised anomaly detection by eliminating the anomaly score thresholding step that is highly dependent on the distribution of the data collected for training. The introduced dimensionality reduction based one-class anomaly detection increases the generalization capability of the trained models during deployment in productive robots. Using logical anomaly detection for detecting pose deviations in the post gripping process has also been considered in this research by introducing a novel vision graph based anomaly detection approach, called ViGLAD. This approach has been shown to outperform state-of-the-art logical anomaly detection approaches in both public datasets and robotic specific datasets, collected in a logistic robotics application. Establishing a workflow to define the steps for developing and integrating state-of-the-art structural and logical image anomaly detection algorithms in productive intelligent robot has been the focus of the final part of this research. The proposed approach, called DevOpsAD, is the process that allows upgrading camera based intelligent robots with anomaly detection based post gripping perception modules without the need to stop operations. A node between development and operation teams is introduced in order to close the gap between robot specialists and data scientists. DevOpsAD is supported by multiple tools for transparency and efficiency, such as the anomaly generator that increases the speed of data collection and trained model evaluation. The impact of efficiently integrating image anomaly detection in object detection based logistics robots has been evaluated based on packaging and a kitting robot applications. The results show that it can substantially increase the quality of the application output by 12% while significantly enhancing the process stability by up to 48% and reducing the down time of the robot per pick and place cycle from 21 seconds to 1 second in average.

Awarding Institution(s)

University of Plymouth

Supervisor

Dena Bazazian, Giovanni Masala, Mario Gianni, Asiya Khan

Document Type

Thesis

Publication Date

2025

Embargo Period

2025-07-26

Deposit Date

July 2025

Creative Commons License

Creative Commons Attribution-NonCommercial 4.0 International License
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License

Download

Share

COinS