Work Tempo Instruction Framework for Balancing Human Workload and Productivity in Repetitive Task

This paper proposes a feedback framework that adjusts human workload and productivity by instructing work tempo and personalizes work according to individual differences. For feedback of optimal work tempo, this study proposes a human worker state transition model to investigate the effects of work tempo instructions on workload and productivity. Based on the results obtained, we proposed a feedback policy using our proposed worker state transition model. By testing our proposed framework in a picking task, we showed the possibility to balance productivity and workload.


INTRODUCTION
Numerous countries are grappling with a diminishing workforce, leading to challenges in securing adequate skilled labor.Previously reliant on a select group of skilled workers, several roles now require the participation of individuals from diverse backgrounds, including temporary staf.Amidst this labor shortage, there is increasing interest in harnessing human-robot interaction (HRI) technologies for versatile support in manufacturing sectors.However, tailoring work to the individual capability of each employee is essential for enhancing worker productivity.
This study devised a framework to monitor, analyze, and provide feedback for work personalization, aiming to balance workload and productivity in the manufacturing industry.Although previous research such as [3] has focused on monitoring workload and dividing tasks between humans and robots, it lacks in providing feedback for work modifcation.Our framework dynamically adjusts task demands to infuence human conditions.Giving appropriate task demands can activate the workers' behavior and may improve their ability to perform the task.Shirakura et al.
[4] demonstrated the efectiveness of using time pressure to manage productivity and workload.
Consequently, we integrated work tempo control into our feedback system.Tempo instruction, known for its efectiveness in tasks such as gait rehabilitation [2], is deemed suitable for the repetitive tasks common in manufacturing.Our approach includes a human worker state transition model that correlates workload with productivity.The system administers work tempo instructions to optimize this balance, considering the worker's condition (Fig. 1).
Experimental results from various work tempo instructions indicate that alterations in work tempo impact both productivity and physical movement, infuencing the workload experienced by workers.Our feedback system was also applied in a pick-and-place task, which is a basic repetitive activity.
Although the ultimate aim was to integrate work tempo control in natural human-robot collaboration, this study initially explored computer-based work tempo instructions as a preliminary step toward robot implementation.

WORKER STATE TRANSITION MODEL
Fig. 2 displays a proposed worker state transition model.In this paper, productivity is defned as the basic production capacity, exemplifed by takt time.Our model considers the rate of achieving work demands as a measure of performance, refecting individual capabilities.We classify worker state into discrete categories based on workload and performance and propose a methodology for determining instructional policies contingent on the current state.
Our hypothesis, based on the proposed model, suggests that workers struggle to adhere to tempo instructions that are either excessively slow or fast (Fig. 3).The primary objective for workers is to maintain the instructed tempo.We employ a latency rate, a = i , as a performance indicator, where a represents the actual operation time and i , the operation time as per the instruction.A latency rate exceeding 1.0 indicates a slower work tempo than instructed, and vice versa.The inability to follow instructions is  categorized as Boreout and Burnout.Additionally, within the manageable tempo range, we identify areas of potential Boredom and Overload.Optimal state is anticipated to lie between these two extremes.Policy decisions must guide workers toward this optimal state.For instance, increasing task demands might be necessary for a bored worker, whereas reducing them could be benefcial for an overload individual.
We hypothesize that optimal tempo instructions can balance the workload and performance of a worker through a two-step process: 1) Adjusting the instructed tempo within a range that the worker can follow.2) Modifying the tempo to ensure the workload is neither too high nor too low.
The current state of the worker is estimated by combining their workload and performance metrics.The present experiments aimed to apply actual worker performance data to this model and validate a feedback system that instructs tempo according to the model's policy.

EXPERIMENT
We conducted two experiments to validate the proposed worker state transition model and the feedback system.

Task
We selected the pick-and-place task, a typical repetitive activity in factory settings, as the focus of our study.This fundamental operation is integral to various tasks.Initially, we evaluated the efectiveness of out proposed framework in a basic scenario before extending to more complex tasks.The experimental setup is depicted in Fig. 4. The participants were required to pick up the parts from a shelf, pushing a trolley by hand, and then place these parts in a box on the trolley.The shelf was positioned to the right of the participant, who used their right hand for the pick-and-place activity while maneuvering the trolley with their left hand.
For this experiment, Lego blocks were selected as the objects to be picked, enabling us to isolate the impact of work tempo variations by eliminating variables such as part shape and weight.The block, in six diferent colors, were stored in distinct boxes on the shelf.The participants were tasked with picking a block and placing it in the specifed section of the parts box on the trolley.
Additionally, a Stroop task [5] was integrated to assess the interplay between workload, performance, and work quality.For example, a blue block should be placed in the top right of the parts box in Fig. 4. Accordingly, the participants were instructed to position the selected item in a location denoted by text on the tray, with each text label being unique and not replicated.The confguration of the parts box and its arrangement were randomly altered for each trial by the experimenter.

Experimental Protocol
Informed consent was obtained from all participants prior to their involvement.The experimental protocol of this the study received approval from the local institutional review board of the National Institute of Advanced Industrial Science and Technology (HF2021-1155).

Experiment 1:
Experiment for the proposed worker state transition model.In this study, 13 individuals (six men and seven women) participated.The average age of the participants was 36.6, with a standard deviation () of 13.7.Initially, each participant engaged in ten practice trials of the pick-and-place task.The objective during training was to perform the tasks swiftly and accurately without error.The base tempo for each participant was determined by averaging the tempo during this training phase.Subsequently, participants undertook the task under varying time pressures ( = 0.9, 1.0, 1.2, and 1.4), with representing a multiplier relative to the base tempo.Tempo instructions were delivered through a metronome sound via headphones.Each condition involved ten trials.The participants were instructed to adhere to the given tempo as closely as possible.
The order of the tempo conditions was randomized for each participant, who were unaware of the tempo and sequence.Post each condition, participants completed a NASA-TLX questionnaire [1].Although questionnaires are a conventional method for workload assessment, they are not feasible for real-time indicator.Therefore, they were used for reference, and joint torque was analyzed as a real-time indicator.Additionally, a minimum break of 10 min was provided between successive trials for each condition and the training sessions.

Experiment 2:
Experiment for the proposed feedback system.Two participants (two men) were involved in this experiment.The feedback policy was established as follows: where denotes the instruction tempo frequency in time step , d = i − h indicates the diference between the instructed tempo and the work tempo frequency of the worker, p refers to the gain that indicates how quickly the instruction tempo changes when the worker cannot follow the instruction, w denotes the gain for the workload, threshold represents the threshold that determines whether a worker can follow the instructions, and threshold indicates the workload when the worker is in the optimal state.In this experiment, p and w were set at 0.01 through trial and error to ensure appropriate tempo changes.We established threshold at 1.2 rad, again by trial and error, to prevent oscillations.The initial instructed tempo, 0 , was based on the base tempo measured for each participant in section 3.2.1.Instruction frequency was updated at 40 Hz.Despite the ideal being real-time indicators such as joint torque, self-reporting was utilized in experiment 2 as a preliminary test for real-time feedback through tempo instructions.The participants indicated task perception by pressing buttons on the trolley: one for boredom, another for neither boredom nor difculty, and no button press for challenging tasks.
This was premised on the assumption that increased task difculty might lead to neglecting button presses.Workload range was set from 0 to 1.0, with optimal workload, threshold , defned as 0.5.The system outputted 0.0, 0.5, and 1.0 as , for bored, optimal, or no button press, respectively.In total, four conditions were tested: without instructions, instructions at a constant low tempo, instructions at a constant high tempo, and dynamic instructions using the proposed framework.

Implementation
Motion capture with OptiTrack, Prime 41, was employed to monitor the full-body movements of the participants.The work progress was assessed using this whole-body motion.The participants wore suits equipped with refective markers, utilizing an Optitrack full body baseline (41) 1 was used as the marker set.Motion data were captured at 100 Hz using using OptiTrack, Motive.After the experiment, the motion data were analyzed with in-house digital human software, DhaibaWorks2 to evaluate the joint torque as an indicator of workload.
The human body model in DhaibaWorks was tailored to each participant's height and weight.A Butterworth flter removed highfrequency noise from the raw motion data.Given the negligible mass of the Lego block 2.5 g, its efect on joint torque was considered minimal, with no external forces assumed apart from ground interaction.

RESULT 4.1 Experiment 1
4.1.1Performance Evaluation.This study analyzed the human conditions under varying tempo instructions, focusing on performance.Fig. 5 illustrates the latency rate as a performance metric in the pickand-place task.The horizontal axis represents the time pressure , whereas the lines depict the latency rate of each participant.We mapped these results onto the proposed worker state transition model, workload, and performance hypothesis (Fig. 3), categorizing them into four classes.
In class 1, latency rates fuctuated with the instructed tempo, refecting states from Boredom to Overload (Fig. 2).Class 2, showed increasing latency rates with faster tempo, indicative of states from  Optimum to Overload (Fig. 2).The consistent latency rates of class 3, irrespective of tempo changes, suggested an alignment with the optimum state of the model (Fig. 2).Class 4, exhibiting nearly linear delay change, appeared to disregard the instructions.Excluding class 4, participants generally adhered to instructions unless the tempo was excessively slow or fast.
4.1.2Workload Evaluation.We assessed human workload through questionnaires and joint torques derived from body movements.A mean correlation coefcient of 0.64 between the instruction tempo and the NASA-TLX WWL score afrmed an increased workload with faster tempos.Table 1 presents the correlation coefcients between joint torque and NASA-TLX WWL scores, focusing on the right shoulder, torso, left hip, and right hip, pertinent to pick-and-place tasks.The coefcients exclude class 4 data from Fig. 5 because of their noncompliance with instructions.
Displayed are the mean and values of Pearson correlation coefcients between the NASA-TLX WWL and the 90-th percentile values of joint torque for all participants.The correlation of the right shoulder torque with NASA-TLX WWL was highest, signifying a strong workload association.

Experiment 2
The results of the experiment are summarized in Table 2.
Operation time refers to the duration between picking successive parts, and workload is measured by the NASA-TLX WWL score.For participant A, instructed conditions yielded higher workloads compared to non-instructed ones.The longest average operation times occurred under constant low tempo instructions, whereas constant high tempo conditions resulted in the shortest operation times and highest workloads.Using the proposed method, the average operation time was second shortest after the constant high tempo, maintaining a workload comparable to constant low tempo.For participant B, the proposed method resulted in operation times similar to low tempo usage and higher workloads subsequent to high tempo conditions.

DISCUSSION AND FUTURE WORK
The performance evaluation outcomes corroborate our hypothesis regarding the worker state transition model: workers struggle to comply with instructions if the tempo is excessively slow or fast.However, the current approach relies on subjective classifcation; for robust validation, a quantitative method such as parametric classifcation is necessary.
Concerning the proposed feedback framework, the results of participant A suggest its efectiveness in balancing performance and workload.Conversely, the outcomes of participant B indicate that the framework might adversely afect the workload and performance of certain workers.
Future eforts will focus on expanding the participant pool to further validate the proposed feedback framework.Additionally, the goal is to develop a feedback framework that minimize reliance on subjectivity, using real-time workload indicators such as joint torque, which have demonstrated correlation with NASA-TLX assessments.

Figure 3 :
Figure 3: Hypothesis of worker state transition model.

Table 1 :
Correlation between NASA-TLX WWL score and 90th percentile of joint torque.

Table 2 :
Results of the experiment with the proposed feedback framework.