TRIADIC MANY-TO-ONE HUMAN-ROBOT COLLABORATION FOR FLEXIBLE MANUFACTURING

Abstract

Collaborative robotics is one of the most promising sectors of robotics, as it enables robots to work in close proximity to humans and work alongside them, through natural communication and safe exchange of mechanical power. In the long term, this is a key enabler for the increased acceptance of robotic technologies in everyday life. Surprisingly, most of the research in human-robot collaboration predominantly considers collaborative interactions involving only one human and one robot in the team composition. Few cases consider interactions with teams composed of a larger number of agents. With the evolution of this field, these interactions are likely to occur more often in the future. Foreseeing this happening, this thesis aims at expanding the research domain of non-dyadic human-robot collaboration. Specifically, collaborations between two humans and one collaborative robot are considered, in the context of manufacturing applications. Research was carried out with the end goal of implementing and validating a robotic architecture capable of autonomously handling a collaborative task with a pair of users working at the same time. Because of the novelty of this research, this thesis proposes a modeling of the non-dyadic interactions that occur in the depicted team dynamics. It revolves around the definition of three states of a single user, that can then be used to infer all possible cases a robotic system can face while interacting with pairs of users. Upon such modeling, this thesis offers a solution for the collection of data regarding these scenarios to train deep learning models to help the robotic system be aware of the ongoing collaboration. In response to the lack of data available for these applications, this thesis shows how it is possible to come up with comparable results in the related multi-user activity recognition problem, starting from data related to single users. This approach was validated by training a stacked long short-term memory and a variational autoencoder. In working out the architecture, this thesis has also conceived a non-dyadic collaborative task, used as testbed to validate the human-robot interactions involved in this novel scenario. It was designed so that the interplay to enact in the collaboration allows users to experience all the possible cases that can occur in the task thanks to the different timings of its sub-steps. The testbed was used in a large-scale user study to understand how a collaborative robot should behave in such a workspace to keep an impartial strategy in supporting both users at the same time. The results achieved in this project led to the development and testing of a robotic architecture, which was successful at autonomously handling the concurrent non-dyadic human-robot collaboration in which it was deployed. The architecture is able to sense the users in the environment. Besides, it computes the pose of an object of interest shared between the users by making use of their positions and orientations in space, their body lengths and their state, following the previously mentioned modeling. In another large-scale user study, the seamless real-time execution of the architecture in the task at hand was deemed acceptable and usable by the users. Both user studies conducted in this research took place at the BAE Systems Academy for Skills and Knowledge Centre, to reach participants involved in manufacturing and collect evaluation from people who are likely to benefit in the future from the envisioned manufacturing applications.

Date of Award	1 May 2025
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Samuele Vinanzi (Supervisor) & Angelo Cangelosi (Supervisor)