**5. Conclusions**

This paper presented the influence of LOA on a human robot collaborative assembly task considering different workload levels. The user study yielded valuable insights into participants' preferences and influence of LOA and workload. The study also introduced two constructs for the evaluation: quality of task (QoT) execution and usability. The evaluation obtained through these constructs highlighted their potential for use in HRI studies. The study has served to provide support tools to further align manufacturing strategies and automation decisions putting into consideration level of workload to further improve productivity.

The QoT execution construct also pointed to the significance of combining efficiency and effectiveness together as a single variable. It revealed the influence of the LOA and workload in the extent to which goal of the task was accomplished under specified degree of accuracy and duration of the task. The usability construct was significant in revealing the combined effect of QoT execution and user perceptions of the ease of use, workload, and system reliability. The interactive effect of LOA and levels of workload on this construct pointed to the added value which user perceptions contribute when combined with the QoT measure.

We recommend a high LOA to support the user when the workload is high. A high LOA could reduce the stress or pressure of additional secondary tasks which the robot could support in. This was observed in the outcome of the user preferences which tended towards higher LOA when the workload was high. It also agrees with the observations of [38] in their meta-analyses considering the influence of LOA on workload. High LOA, when designed effectively, helps to extend the capabilities of the user to attend to other tasks concurrently as noted by [42,61]. However, lower LOA is helpful when high task complexities are involved, for which failure performance may occur as also noted in [39]. An adaptive LOA design that takes these outcomes into consideration is therefore recommended for further investigation.

There may be significant differences in the influence of these variables when observed in other settings, with different forms of robots, tasks and robot feedback modalities [62] and with the perception of different users as emphasized in [63]. Future work should evaluate different forms of increased workload. The workload design can be fine-tuned to portray distinct types of workload demands such as physical, cognitive and temporal demands during the task. Evaluation should also be conducted with other forms of tasks e.g., with a mobile robot delivering items and with other populations. Ongoing research is aimed at performing studies with older adults for daily living tasks and for nonprofessional users, putting into consideration the influence of demographics on the changes automation brings [64]. LOA has proven to influence performance for older adults [12]. We expect the effect of the levels of workload to amplify with them. The change of preferences and the differences in the reaction and performance of the older adults should be examined with different LOA options for different workload levels.

**Author Contributions:** Conceptualization, D.G., S.O. and Y.E.; methodology, D.G., S.O. and Y.E.; software, D.G.; validation, D.G., S.O. and Y.E.; formal analysis, D.G., S.O. and Y.E.; investigation, D.G., S.O. and Y.E.; resources, Y.E.; data curation, D.G.; writing—original draft preparation, D.G.; writing—review and editing, D.G., S.O. and Y.E.; visualization, D.G., S.O. and Y.E.; supervision, S.O. and Y.E.; project administration, Y.E.; funding acquisition, Y.E. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was supported by the EU funded Innovative Training Network (ITN) in the Marie Skłodowska-Curie People Programme (Horizon2020): SOCRATES (Social Cognitive Robotics in a European Society training research network), grant agreement number 721619. Partial support was provided by Ben-Gurion University of the Negev through the Agricultural, Biological and Cognitive Robotics Initiative, the Marcus Endowment Fund, and the W. Gunther Plaut Chair in Manufacturing Engineering.

**Institutional Review Board Statement:** This study was approved by the ethical committee of the Department of Industrial Engineering and Management at Ben-Gurion University of the Negev.

**Informed Consent Statement:** Informed consent was obtained from all subjects involved in the study.

**Data Availability Statement:** Data supporting the reported results can be found at https://github. com/samuelolatunji/LOA-WorkloadLevels\_Analyses.git (accessed on 9 August 2021).

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


**David Ortega-Aranda 1,\*, Julio Fernando Jimenez-Vielma 1, Baidya Nath Saha <sup>2</sup> and Ismael Lopez-Juarez <sup>3</sup>**


**Abstract:** Assembly tasks executed by a robot have been studied broadly. Robot assembly applications in industry are achievable by a well-structured environment, where the parts to be assembled are located in the working space by fixtures. Recent changes in manufacturing requirements, due to unpredictable demanded products, push the factories to seek new smart solutions that can autonomously recover from failure conditions. In this way, new dual arm robot systems have been studied to design and explore applications based on its dexterity. It promises the possibility to get rid of fixtures in assembly tasks, but using less fixtures increases the uncertainty on the location of the components in the working space. It also increases the possibility of collisions during the assembly sequence. Under these considerations, adding perception such as force/torque sensors have been done to produce useful data to perform control actions. Unfortunately, the interaction forces between mating parts produced non-linear behavior. Consequently, machine learning algorithms have been considered an alternative tool to avoid the non-linearity. In this work we introduce an assembly strategy for an industrial dual arm robot based on the combination of a discrete event controller and Deep Neural Networks (DNN) to solve the peg-in-hole assembly. Our results show the difference between the use of DNN with one and with two force/torque sensors during the assembly task and demonstrate a 30% increase in the assembly success ratio when using a double force/torque sensor.

**Keywords:** robotic assembly; deep neural networks; peg-in-hole; dual-arm
