Enhancing Human-Robot Collaboration through Gaze-Based Turn-Taking in Kitting Scenarios

The world of industrial robotics has seen a rapid evolution for the last decade, adapting to the requirements of a constantly advancing manufacturing industry. One of the recent advancements this field has seen is the introduction of collaborative robots into production lines; however, this “collaboration” is often reduced to the human worker adapting and helping the robot perform a pre-defined task. This thesis aims to improve Human-Robot Collaboration by implementing a system which allows the robot to react in real-time to the intentions of a human operator. For this, a system built around the concept of Proactive Eye-Gaze had to be designed and tested. This involved the development and implementation of two Deep Learning-based gaze estimation methods (feature-based and appearance-based), the integration with a collaborative robot (UR- 10e), testing the real-time performance and accuracy of both models in controlled environments and conducting a survey with participants to evaluate said system. The results of the model tests show that the feature-based model has better computational efficiency and, comparing the precision of both models, the feature-based one also outperformed the appearance-based model. The survey results proved that Proactive Eye-Gaze can significantly improve Human-Robot Collaboration, and that replacing wearable gaze tracking devices is possible while maintaining a high level of precision in gaze estimation.