Manual assembly learning, disability, and instructions: an industrial experiment

Abstract

Cognitive assistance systems help people with learning disabilities to increase their skills and consequently their employment opportunities in the regular labour market. Research on advanced work instructions has encouraged training disabled workers in cognitively demanding production tasks, especially manual assembly. However, studies lack evidence on the effect of repetition or work cycle alongside the form of instruction and type of disability. This paper addresses this gap and reports on an experiment conducted at a sheltered workplace. Four forms of instruction (paper-based, animations, projection, adaptive projection) were tested to assist operators with three types of disability (illiterate, psychosocial, cognitive) with a manual assembly task. The results show that projection enhances the first assembly cycle. Challenging operators by filtering the content of the instruction with increased experience leads to greater independence and a better understanding of their tasks. However, adaptive instructions can form a barrier for those operators who are most dependent on mentor support. The form of instruction should thus be considered carefully for each operator as their adaptation to changes and cognitive assistance systems varies. The results are discussed in light of the Industry 5.0 human-centric and socially sustainable production agenda with managerial and research implications and future research priorities.

KEYWORDS:

• Disability
• instruction
• cognitive
• assistance systems
• learning
• Industry 5.0

1. Introduction

The European manufacturing industry has a crucial role in both economic and societal prosperity as it provides 15% of GDP while employing about 35 million people (Statista 2020; Ivanov 2023). The sustainability gap resulting from the demographic transition puts pressure on increasing employment rates and recruiting from a diverse population. However, manufacturing tasks, such as manual assembly, change often and are complex and, thus, require continuous learning. Learning in labour-intensive manufacturing is associated with gaining experience through increasing cumulative production or repetitions (Peltokorpi and Niemi 2019; Glock et al. 2019). Thanks to modern technologies, workers are provided with digital assistance systems to support learning (Hoedt et al. 2017; Mark, Rauch, and Matt 2022). The recently-released European ‘Industry 5.0’ aims at more human-centric, socially sustainable, and resilient production (Bai, Zhou, and Sarkis 2023), where the idea is to choose technologies based on an ethical rationale to support human values and needs (Longo, Padovano, and Umbrello 2020). Such a transition focuses on the interconnection between human intelligence and cognitive computing to assist with a human’s physical, sensorial, and cognitive capabilities. For example, a cognitive assistance system (CAS) (Mark, Rauch, and Matt 2022) can use artificial intelligence (AI) and augmented reality (AR) as enabling technologies. In the Industry 5.0 concept, these technologies are in the category of individualised human-machine interaction (Longo, Padovano, and Umbrello 2020).

CASs have mainly been tested with regular industrial workers (Mark et al. 2019; Schein and Rauschnabel 2022) but should be adapted to diverse groups of users (Mark, Rauch, and Matt 2022) with special needs, e.g. older employees, immigrants, and those with learning disabilities (LDs). LDs refer to ‘a heterogeneous group of disorders manifested by significant difficulties in the acquisition and use of listening, speaking, reading, writing, reasoning, or mathematical abilities’ (Seok, Meyen, and DaCosta 2010). People with disabilities are often outside the labour market; for example, in the EU area, only 51% of them are employed (European Commission 2021). In the United States, only 33% of working-age people with disabilities are employed, and two-thirds of the unemployed ones are willing to work but cannot find a job (Lindsay et al. 2018).

The design of CAS in terms of work instructions comprises two important aspects, which are the form of the information (the carrier or the means of storing information) and the content (the actual information) that is being conveyed to the operator (Fässberg, Fasth, and Stahre 2012). Our review of assembly experiments on the instruction and disability effect in Section 2.2 shows that the form of the information has dominated the content. Those studies often compare paper-based to various digital instruction forms, such as screen- (or monitor-) based and projection (which uses AR), and, less frequently, the content presented within a single form. Schein and Rauschnabel (2022) point out that studies typically highlight the positive impacts of AR, with little attention being paid to the human aspect and potential barriers of AR as a workplace technology. In this line of research, one of the latest advances is the adaptation of instructions to individual needs, providing content that is relevant to the current skill level of a worker (Petzoldt et al. 2020).

There has been a growing interest in the potential benefits of CAS at assembly tasks for subjects with various, mainly cognitive, disabilities (Korn, Schmidt, and Hörz 2013; Funk, Mayer, and Schmidt 2015; Bosch et al. 2020; Vanneste et al. 2020). These studies report on the average performance, but very rarely on the improvement of performance, or learning, over repetitions, which is constantly needed in today’s industry. In addition, Mark et al. (2019) recently showed that the current literature lacks attention to (1) the practical implementation of assistance systems in the industry for people with disabilities, and (2) the allocation of assistance systems to different types and levels of disabilities. Despite the acknowledged importance of CAS in supporting the inclusion of people with LDs in production, a question remains to be answered: How does the CAS, in more practical terms, the form of instruction, affect the assembly learning curve for people with specific disabilities?

This paper contributes to closing this gap by experimentally studying the effect of the form of instruction (paper-based, animations, projection, adaptive projection) and the type of disability (illiterate, psychosocial, cognitive) on industrial assembly performance. The effect on assembly cycle time (CT) is tested during three phases of the learning curve: the first cycle, when learning ceases (plateaus), and in the retention (plateau) phase. The ultimate goal of this study is to find out the potential of projection, specifically its adaptive version, in supporting the inclusion of people with various LDs in assembly production.

The paper is organised as follows. Section 2 provides the related literature and hypotheses. Section 3 describes the methodology used to conduct the experiment. Section 4 presents the results and Section 5 the discussion, with Industry 5.0 managerial and research implications, limitations, and recommendations for further research. Finally, Section 6 concludes the research.

2. Related literature and hypotheses

2.1. Cognitive load and learning

When a human performs an activity, a memory trace is formed. It aids a person in recalling techniques and procedures performed previously. Repetitions (work cycles) reinforce learning and improve performance as recall from memory becomes faster and dexterity increases. Cognitive load theory (CLT) (Sweller 1988) is concerned with how cognitive resources are distributed during learning and problem-solving. CLT distinguishes three types of loads: extraneous, intrinsic, and germane. An extraneous, or irrelevant, load interferes with the learning task at hand (Clark, Nguyen, and Sweller 2011). This could, for instance, be a conversation in the background while an operator is focusing on an assembly task or just work instructions that feature too much text. An intrinsic load is caused by the natural complexity of the task or the concept and information being presented (Sweller 2011). The intrinsic load is proportional to the interactivity of elements in the learning material. An assembly drawing that is composed of many interconnected elements is, for example, intrinsically difficult to understand (Peltokorpi and Niemi 2019; Peltokorpi and Jaber 2022). Unlike an extraneous load, an intrinsic load cannot be eliminated (as material needs to be learned); however, it can be minimised by providing the information in a structured and comprehensive way. Extraneous and intrinsic loads are additive and determine the total cognitive load. The germane load relates to schema acquisition (chunking), which is further directed to the interacting elements associated with the intrinsic cognitive load (Sweller 2011). For example, a schema could be a mental structure of a subassembly that needs to be processed in working memory, without the need for processing all the interacting elements of that subassembly. The learning capacity freed by reducing the extraneous load should be used to promote the germane load, to the limits of the capacity of working memory, which enhances learning. Avoiding the extraneous, managing the intrinsic, and promoting the germane load depends on the characteristics of the subject (e.g. level of experience, disability), the complexity of the task, work conditions, instructional methods, and their mutual relations (Kirschner 2002).

2.1.1. Learning curve

Production and operations management (POM) research uses learning curves (LC) to measure human performance when a psychomotor task is performed repetitively (Glock et al. 2019). The simple power-form LC of Wright (1936) is a standard on which further modifications are built. Wright’s learning curve, henceforth WLC, is of the form: (1) $y_n=y_1{n}^{-b}$(1) where $y_n$ and $y_1$ are the execution times for the $n$th and the first cycle, respectively, and b = −log(∅)/log(2), 0 ≤ b ≤ 1, is the learning exponent, where ∅ = % is the learning rate divided by 100. When fitting a trend line to raw data, the plot usually shows a considerable scatter around the trend line. The scatter in the data is caused by many factors, and their mutual relationships, which are interconnected with the cognitive load. Jaber et al. (2021) recently adjusted WLC (Eq. 1) for cognitive interference by assuming that an operator cannot retain all the information from cycles, but some information is lost over time. Their interference-adjusted WLC, henceforth IAWLC, is of the form: (2) $y_n=y_1{n_e}^{-b}=y_1{\left(\frac{1-e^{-\mathrm{\gamma n}}}{e^{\gamma }-1}\right)}^{-b}$(2) where $n_e$ is the cumulative residual experience (after interference), $\gamma =\mathrm{\alpha t}$ is a fitting parameter, $\alpha$ is a decay exponent, 0 ≤ $n_e,\gamma ,\alpha$ ≤ 1, and $t$ is a fixed time step in the calculation procedure. By increasing the decay exponent (indicating greater interference), the estimated curvature and level at which learning plateaus increase. When interference is not captured ($\gamma$ = 0), the model behaves like WLC. IAWLC was found to outperform other LCs, as it captures characteristics that are related to assembly processes better, i.e. it predicts poor performance during the first cycle (which is closely related to cognitive interference), followed by rapid learning in subsequent ones. Adjustment for cognitive interference improved the curve fit to assembly learning data from subjects who have cognitive disabilities (Jaber et al. 2021). The recent modification of IAWLC with a composite of cognitive and motor LCs (Peltokorpi and Jaber 2022) fitted various learning profiles, including ‘early strugglers’ and ‘early adopters’, well. Note that learning may cease for many reasons. In psychomotor tasks, this happens because cognitive task elements dominate over motor elements during earlier cycles, while motor elements (which are subject to a minimum or standard time) dominate over cognitive elements during later cycles. As workers gain experience, variation in performance typically decreases.

2.2. Learning with disabilities, using instructions

Forms of instruction should consider well-defined principles for handling various types of cognitive load (Annex A). These principles hold even more weight for those with LDs. Disabled learners may suffer from working memory deficits related to a phonological loop that is closely linked to the complex cognition required during many industrial tasks (Swanson and Siegel 2001). In general, tasks that place high demands on cognitive processing are difficult for those with LDs, as they have trouble with attentional processes (Mark, Rauch, and Matt 2022), especially while facing distraction. Problems with metacognitive regulation (i.e. how we control cognition) result in an individual inability to actively engage in learning the task at hand (Seok, Meyen, and DaCosta 2010).

Mark et al. (2019) found that cognitively disabled persons may benefit from software systems, projection-based support, and AR in general. The design principles that aim at managing the intrinsic load (pre-training, segmenting, and allowing learners to control the pacing) are particularly beneficial for them (Seok, Meyen, and DaCosta 2010). The promotion of meaningful learning (memorising and deep understanding of instructional material) can be induced by multimedia learning, which integrates text and pictures (Mayer 2011). Pictorial instructions especially are helpful for those with impaired communication skills. Mark, Rauch, and Matt (2022) recently proposed a methodology for the selection of an assistance system based on parameters that take into account the workplace, the type of disability, and the particular work task. The methodology was tested for mentally disabled workers who assembled components of a door lock at a sheltered workplace. The mental limitation was reflected in a lack of concentration, spatial imagination, and creativity. In this case, projection-based CAS best matched the conditions.

Most importantly, the form of the instruction should be driven by the learner as those with LDs also interact with CAS in different ways than regular persons do (Seok, Meyen, and DaCosta 2010). For example, extensive sound and visual projection elements can be disturbing for users who are sensitive to sensorial stimuli (Grund et al. 2020). On the other hand, additional mentor efforts may be needed during technology-assisted training (Bosch et al. 2020).

2.2.1. Case studies

Table 1 gives an overview of thirteen case studies on the effect of the form of instruction and disability on assembly performance. For each contribution, the number of participants (N) and the type of their disability, the use case (whether industrial or not), the principle of pre-training, and the ranking of instruction forms (or disability types) per dependent variable (mainly CT and error rate) are presented. Among the different forms of instruction, the most elementary one is paper-based and that served as a baseline. Unless otherwise mentioned, the instructions consisted of the same content, i.e. text and images, presented step by step. Screen-based instructions are very similar to paper-based ones but are shown on a monitor. This often helps the operator as there is no need to search for the right instruction (page) after each step. By pushing a button, the next step is supported with the right instructions. Projection is a common way of showing AR instructions that uses camera-based technology to display crucial information in the operator’s field of view. Projection is the only form among AR technologies that does not require a see-through display, such as smart glasses, tablets, etc.

Table 1. Overview of empirical studies on the effect of the form of instruction and disability on assembly performance (n/a = significance not assessed, * = significant difference, otherwise non-significant difference).

#	Contribution	N	Type of disability	Industrial use case?	Pre-training	Ranking of instruction (or disability) per dependent variable
1	Furniss et al. (1999) (1)	1	Mental (Vineland Scale = 15%)	Yes	Two cycles w/o support	CT (n/a)
						1. Screen-based with audio prompts, 2. w/o prompts
						Error rate (n/a)
						1. with prompts, 2. w/o prompts
2	Furniss et al. (1999) (2)	1	Mental (Vineland Scale = 40%)	Yes	Two cycles w/o support	CT (n/a)
						1. Screen-based, 2. Paper-based
						Error rate (n/a):
						1. Paper-based, 2. Screen-based
3	Korn, Schmidt, and Hörz (2013)	40	Cognitive and/or physical	No	n/a	CT:
						1. Projection, 2. Screen-based
						Error rate:
						1. Screen-based, 2. Projection
4	Lancioni et al. (2013)	2	Intellectual (profound) disability and total blindness	No	One cycle with mentor guidance	Completion rate (n/a):
						1. Audio prompts, 2. Human assistance
						Error rate (n/a):
						1. Audio prompts, 2. Human assistance
5	Lancioni et al. (2014)	3	Intellectual (profound) and visual impairment, and deaf	No	One cycle with mentor guidance	Completion rate of correct assembly (n/a):
						1. Light cues and stimulation device
						2. Human assistance
6	Funk et al. (2015)	15	Cognitive (PI = 5–10%, 15–35%, or 40%+)	No	n/a	Time per part*:
						1. Projection, 2. Screen-based
						Errors per part*:
						1. Projection, 2. Screen-based
7	Funk, Mayer, and Schmidt (2015)	62	Mental (PI = 5–10%, 15–35%, or 40%+)	Yes	One cycle while instructed verbally	CT:
						1. Contour, 2. Pictorial, 3. No visualisation, 4. Video
						(Disability affected negatively*) Error rate:
						1. Contour, 2. Video, 3. Pictorial, 4. No visualisation
						(Disability affected negatively*)
8	Korn, Funk, and Schmidt (2015)	10	Mild cognitive, and partly mild physical	Yes	No, subjects were familiar with product	CT:
						1. Projection, 2. Screen-based
						Error rate:
						1. Screen-based, 2. Projection
9	Korn et al. (2016)	5	Mental (PI = 5%, 10%, 20%, 20%, and 35%)	Yes	30 min of assembly	CT:
						1. Projection (with gamification)
						2. Screen-based
						(* for PI = 35%, for PI = 5% opposite rank)
10	Kosch et al. (2018)	15	Cognitive (PI = 5–10%, 20–35%, or 40%+)	No	One cycle with paper instruction	CT:
						1. Projected*, 2. Auditory, 3. Haptic*
						Error rate:
						1. Projected, 2. Auditory, 3. Haptic
11	(Bosch et al. 2020) (1)	17	One or more unspecified disabilities per worker	Yes	n/a	CT at the beginning:
						1. Human assistance only, 2. Human and projection
						CT at the end:
						1. Human and projection, 2. Human only
12	(Bosch et al. 2020) (2)	19	Unspecified disabilities per worker	Yes	n/a	Learning speed (n/a): 1. Human and projection, 2. Human only
13	(Vanneste et al. 2020)	44	Cognitive (various levels or non-disabled)	Yes	One demo cycle with oral explanation	CT:
						1. Paper-based, 2. Oral, 3. Projection
						Error rate:
						1. Projection*, 2. Paper-based, 3. Oral
						Independence:
						1. Projection*, 2. Paper-based, 3. Oral*
						(Disability affected all parameters negatively)
	This study	24	Illiterate, psychosocial, cognitive	Yes	No	Paper-based, animations, projection, and adaptive projection are compared in terms of
						CT in the first cycle
						CT when learning plateaus
						Retention in the plateau phase

Furniss et al. (1999) experimented with whether two mentally disabled subjects could assemble a valve faster and with fewer errors by using a screen-based handheld device. A severely disabled subject performed better on both measures when audio prompts in the device were enabled. A less disabled subject was guided by paper- and screen-based instructions. The latter condition resulted in faster performance, while the paper instructions produced slightly fewer errors.

In the study of Lancioni et al. (2013), two totally blind and profoundly intellectually disabled subjects assembled five-component trolley wheels. The baseline condition included, when necessary, only verbal assistance by an experimenter. An intervention with technology provided prompts, spatial guidance, and positive feedback in a way that enabled a faster task completion rate and the prevention of errors. Lancioni et al. (2014) conducted another study with three multi-disabled subjects (each deaf and visually and intellectually impaired) who assembled five-component water pipes. Human-assisted conditions were compared with an intervention using light cues and a stimulation device. With the help of technology, the subjects were able to perform the assembly consistently and independently. In the old conditions, pipes were assembled correctly at a very slow rate while continuous human assistance was needed.

Korn, Schmidt, and Hörz (2013) compared screen-based instructions and in situ projection (without feedback on errors) for cognitively and/or physically disabled subjects whose instructions were normally paper-based. The subjects assembled eight Lego cars, each consisting of nine bricks. The use of projection, on average, reduced CT by 8%, while the absolute increase in the error rate was 6.1%. In general, projection was seen as a way to reduce cognitive load. However, a more detailed analysis showed that 40% of those receiving instructions through projection performed worse than usual, and in particular there was a lack of quality. The authors concluded that the use of in situ projection should be restricted to those whom it suits well. For the ones it did not suit, they provided a potential explanation: they are slow learners and need more time to get adjusted to the new technology.

In the study of Korn, Funk, and Schmidt (2015) mild cognitively (and partly physically) disabled subjects assembled metal shears in a process consisting of five steps. When there was a change from screen-based instructions to projection (with gamification), they were, on average, 6.4% faster but with a 6.0% absolute increase in their error rate. As the projection conditions in Korn, Schmidt, and Hörz (2013), and Korn, Funk, and Schmidt (2015) did not involve quality feedback (real-time error detection), a speed-accuracy trade-off was likely to occur. Korn et al. (2016) studied the forms of instruction and assembly, like Korn, Funk, and Schmidt (2015), but with five mentally disabled subjects during an entire workday. Each subject was assigned to one of the disability categories based on his/her performance index (PI), indicating a percentage capacity that the subject can perform a task against a regular worker. The most capable (PI = 35%) subject was significantly faster with projection and the least capable (PI = 5%) with screen-based instructions.

Similarly, Funk et al. (2015) studied the assembly of a Lego product with cognitively disabled workers assigned to screen-based instructions or in situ projection (with feedback on errors). They found that, by using projection, the workers could assemble more complex products up to three times faster and with up to 50% fewer errors. In another study, Funk, Mayer, and Schmidt (2015) compared pictorial, video, and contour visualisations for projected instructions to assemble a clamp consisting of five parts. 62 mentally disabled subjects participated in the experiment. The group using contour visualisation was the fastest and made the fewest errors, pictorial visualisation was the second fastest, and the group with video visualisations made the second fewest errors.

Kosch et al. (2018) compared three different error-feedback modalities, visual (projected light), auditory (error sound), and haptic (vibrating gloves), for cognitively impaired subjects who assembled a Lego product using paper instructions. Each feedback modality was used once per subject. Visual feedback outperformed auditory and haptic feedback in terms of both productivity and quality.

Bosch et al. (2020) reported two case studies where novices were trained by a human or, in addition, projection guidance. In both studies, the subjects had one or more unspecified disabilities. In the first study, a part of a solar panel was assembled. At the beginning of the training session, those supported only by a human were 45% faster. In the end, the use of projection guidance, where the amount of information was reduced every 10 min, was 22% faster. It is noteworthy that additional human support was needed with the use of projection. After training, in the retention phase, the projection support was removed and no significant difference in CT was found between the subject groups. For both groups, human support decreased by 90% from that at the beginning of the training. The second study investigated the amount of training needed to reach a predetermined CT in a light assembly context. In the old situation, with only human support, it took one to eight hours. With projection guidance, three cycles were needed for, on average, a 30% reduction in CT, and two-thirds of the subjects needed less human support.

Vanneste et al. (2020) studied three tasks, assemblies of a light fixture and wheel and quality control of an aluminium part, with subjects with varying levels of cognitive disability and experience. Each subject first performed one of the three assembly tasks four times, and then the two remaining tasks once each. The capabilities of the subjects and the tasks and the sequence in which they were performed were counterbalanced between three forms of instruction, oral, paper-based, and projection. The results show that projection slightly increases the mean CT in comparison to paper and oral instructions. The advantage of better productivity with projection in the first cycle disappeared during the subsequent cycles, probably as a result of the operators being only distracted by the projection while needing less support with increasing experience. Projection reduces the mentor’s effort during training and allows the operator to access more complex assembly activities.

In only one case study in Table 1 (#11, [Bosch et al. 2020]) projection support was steadily reduced as the subjects gained more experience. Hence, the next section reviews the literature on adaptive instructions, which are also used in our experimental study.

2.2.2. Adaptive instructions

The effect of individual adaptivity in AR instructions has, to our knowledge, not yet been studied with disabled subjects (Heinz et al. 2021). Petzoldt et al. (2020) studied a CAS comprising four functionalities: (1) instruction system; (2) quality and completeness control; (3) progress recognition; (4) individualisation of support and worker qualification. They stated that the assistance system should adapt to the needs of an individual worker, provide only relevant information, and decide the appropriate detailing of instructions. This would happen by measuring appropriate indicators to calculate a corresponding qualification score. As for performance indicators, they propose the average throughput time for a given number of assembled products and its variation. This would also apply to individual process steps and searching time. Furthermore, they suggest this approach to facilitate the training of new employees.

Kelleher and Hnin (2019) conducted experiments to explore and provide insights into the effect of the average number of attempts, steps, and restarts on the degree of struggling (the rated cognitive load) when completing puzzles. They also developed a model with the intention of predicting the cognitive load on a future code puzzle problem. For future work, they recommend developing adaptive learning systems that shape the learning experience in new ways, such as predicting the cognitive load from learning curves.

Tsutsumi et al. (2020) used multiple physical human properties and a self-evaluated assembly skill level to estimate the experience level of a worker before the start of production. Along the process, the experience level is updated after a certain number of assembly iterations. They state that changing the instructions should be well-considered since it could cause the workers some confusion. Heinz et al. (2021) presented a concept and implemented an assembly line with interactive in situ projections to detect changes in the performance levels based on the analysis of the processing times of workers with and without disabilities. For further studies, they suggest investigating the ways in which ‘workers with disabilities can benefit from performance-based adaptations’.

2.3. Hypotheses

The LC effect received very little attention in the case studies (Table 1), despite its importance in the POM field and manual assembly (Glock et al. 2019; Peltokorpi and Jaber 2022). In this study, the effects of the form of instruction and type of disability on assembly CT are tested during three phases of the LC: the first cycle, when learning ceases (plateaus), and in the retention (plateau) phase. The plateau CT, or threshold, is defined later in Eq. (5) in Section 4.1.2. We assume, like Jaber et al. (2021), that an operator cannot retain all the information from cycles, but some information is lost over time. We fit their LC, in Eq. (2), to individual learning data, to estimate the level of cognitive interference and retention.

The following six research questions (RQ) and hypotheses (H) are used. Hypotheses H1-H3 are for the effect of instruction form and are derived from the case studies. Instead, H4-H6 are derived from the experience of the present sheltered workplace, as the effect of disability type has not been studied before.

RQ1: How does the form of instruction affect the first assembly CT for disabled workers?

H1: More advanced instructions have a positive impact on the first assembly CT. The mean of the first assembly CTs is the shortest with projected instructions, followed by animations and paper instructions.

RQ2: How does the form of instruction affect CT when learning plateaus?

H2: More advanced instructions have a positive impact on CT when learning plateaus. After a percentage improvement, the mean of the CTs at which learning plateaus is the shortest for adaptive projection, followed by projection, animations, and paper instructions.

RQ3: How does the form of instruction affect retention in the plateau phase?

H3: In the plateau phase, instructions are no longer needed, and workers perform equally. Retention in the plateau phase in terms of the percentage of CTs below the threshold and the level of cognitive interference is the same with each form of instruction.

RQ4: How does the type of disability affect the first assembly CT?

H4: Illiteracy slows down the comprehension of new learning material. The mean of the first assembly CTs is longer for illiterate subjects than for psychosocially and cognitively disabled ones.

RQ5: How does the type of disability affect CT when learning plateaus?

H5: Cognitive disability has a negative impact on CT when learning plateaus. After a percentage improvement, the mean of the CTs at which learning plateaus is the longest for cognitively disabled subjects, followed by psychosocially disabled and illiterate subjects.

RQ6: How does the type of disability affect retention in the plateau phase?

H6: Retention in the plateau phase is the worst (the percentage of CTs below the threshold is the lowest and the level of cognitive interference is highest) for cognitively disabled subjects, followed by psychosocially disabled and illiterate subjects.

3. Methodology

3.1. Case product and participants

To investigate the hypotheses, industrial experiments were conducted at Senzer, a sheltered workplace located in the Netherlands, where people with various disabilities are given the opportunity to be employed. The case product was the adjuster of a Dorel’s Maxi Cosi car seat. This seat adjuster consists of three subassemblies and a total of 17 parts to be assembled.

A total of 24 subjects (both male and female) participated in the experiment and were divided into three groups (with eight subjects in each) based on their disability. Illiterate (ILLIT) cannot read, psychosocially (PSYSOC) disabled subjects are either sensitive to stress or more dependent on mentor support, and cognitive (COG) disabled subjects have intellectual issues or are just slow learners.

3.2. Instructions

Four forms of instruction were studied: paper-based (PAPER), animations (ANIM), projection (PROJ), and adaptive projection (ADAPT). The elements used for each form of instruction are presented in Figure 1, inside the corresponding box. In general, the instructions follow the design principles and guidelines presented in Annex A. Each form of instruction shows the required tools and parts and the progression after each of the 49 micro steps to complete the three subassemblies of a seat adjuster.

Figure 1. Forms of instruction studied with the elements inside a box.

At heart of the figure is paper instructions’ element box including text and image. Animations add GIFs, annotations, and next step. Projection adds guide light. Adaptive projection shows or hides these elements based on the experience level.

Paper-based instructions, with text and images, form the baseline for more advanced instructions. Animations present text and images on a screen extended with GIFs (where possible) and annotations, such as arrows and wrong/right symbols. The operator controls the transition to the next step with a button. The projection extends the animations with a blinking guiding light, and all the instructional elements are projected on the tabletop using the HIM¹ (Human Interface Mate) spatial AR technology. The HIM is a commercially available device, developed by Arkite,² that tracks and registers the progress of manual assembly operators via a vision and infrared sensor system. The HIM uses AI technology to automatically detect when a step is finished and shows the next step. Thus, the HIM fits perfectly into the Industry 5.0 concept.

Adaptive projection uses an algorithm to determine the operator’s experience level, with 1 indicating a beginner (equals ‘projection’) and 4 the most experienced level (as in Moghaddam et al. (2021)). More precisely, the experience level is determined for 17 step groups, each consisting of an average of three sequential micro steps, for example: (1) grab the axle pusher, (2) push the axle against a stop, and (3) put the axle pusher back. Figure 2 depicts an example of the adaptation of an operator’s experience level as a function of assembly cycles for this step group. To determine the current experience level, the algorithm considers the CTs of the last three step groups and makes a comparison with the step group CT of an outstanding operator. In this example, the operator reaches an experience level of 3. When the experience level increases, the number of instructional elements shown (within the step group) decreases, and vice versa. Level 2 (column 2 in Figure 1) hides text and GIFs, level 3 images, and level 4 annotations. At levels 3 and 4, the guiding light is set to ‘solid on’ as the operators are more familiar with the locations and a blinking light would only distract them.

Figure 2. Example of operator experience level (adaptivity of instructions) based on algorithm.

In this scatter chart, each point represents step group cycle time from cycle 1 to 33. A trend line illustrates how operator’s experience level improves from level 1 to 3 through cycles.

3.3. Experimental procedure

The study used a between-groups (types of disability, forms of instruction) repeated measures (numbers of assembly cycle) design with random assignment of subjects to the instruction forms. The experiments took four weeks, with one instructional group per week. Each day, two subjects executed the assembly of a seat adjuster as many times as possible during a two-hour session. The subjects had no prior experience with this product or the HIM unit, and they were not allowed to conduct pre-training. A short introduction to the assembly task was given by a mentor, along with a brief explanation of the form of instruction. The mentor helped with any substantial difficulties to ensure the progression of the operator. This happened at the initiative of either the operator or the mentor. The first runs were video-recorded as the subjects were most probably struggling with the tasks and might need substantial help from the mentor. This made it possible for the researcher in the later analysis to count the times each subject was helped. At the end of the experimental session, the mentor could write a general comment about how the subject performed. Note that mentor help and comments are not considered primary indicators but provide only additional insights into the performance. After cycles 1, 2, 4, and 8, the subjects had a break of one minute.

4. Results

For hypotheses H1-H6, the effects of the form of instruction and type of disability on initial and plateau CT and retention in the plateau phase were analysed. The analysis was conducted using the IBM SPSS Statistics v.28 software. The mean CTs at instruction form level and disability type level in each cycle are calculated as follows: (3) ${\overline{\mathrm{CT}}}_{\mathrm{ik}}=\sum _{n}C{T}_{\mathrm{ikn}}/N_{\mathrm{ik}}\text{for instruction form}i\text{,}\text{at}\text{cycle}k$(3) (4) ${\overline{\mathrm{CT}}}_{\mathrm{jk}}=\sum _{n}C{T}_{\mathrm{jkn}}/N_{\mathrm{jk}}\text{for}\text{disability}\text{type}j,\text{at}\text{cycle}k$(4) where $N_{\mathrm{ik}}$ and $N_{\mathrm{jk}}$ are the total numbers of subjects with the form of instruction $i$ and type of disability $j$ in cycle $k$. In most cases, 15 cycles were considered. Some subjects did not reach this number of cycles, which has been taken into account in the results and analysis.

4.1. Instruction effect

Figure 3 presents the mean CT for each form of instruction and cycle (Eq. 3).

Figure 3. Effect of form of instruction on mean and SD (2σ) of CT in each cycle.

Paper instructions and animations result in longer first cycle time and steeper initial learning curve. Each curve levels out until cycle 15. With paper instructions, within-group variance is very high in the first cycle and with adaptive projection later.

4.1.1. First cycle time

The mean of the first assembly CTs is the shortest with projection-based (PROJ and ADAPT) instructions. Since the first cycle is not subjected to adaptivity, these groups were combined (with 885.8 ± 207.9 sec). ANIM (1182.5 ± 292.7) and PAPER (1178.4 ± 665.8) performed worse in this regard. According to an independent samples T-test, with ANIM, it took significantly longer in the first cycle than with projected instructions (p < 0.012; size effect Cohen’s d = 1.249). Between PAPER and projected instructions, the difference is non-significant (p < 0.086; d = 0.714). Thus, H1 is partially validated. In the first cycle, the mentor helped significantly more often those using projections (7.75 ± 3.39 times) than PAPER (3.83 ± 4.07; p < 0.023; d = −1.083) and ANIM (4.33 ± 3.14; p < 0.028; d = −1.031).

4.1.2. Plateau cycle time

After a rapid initial learning phase, LC levels out i.e. enters a plateau. Entering the plateau, or steady-state, phase needs sufficient training (cycles). Therefore, the plateau analyses in this study exclude the six operators who finished less than 14 products, i.e. one operator with COG and one with PSYSOC disability in the ANIM group, and two COG disabled in both the PROJ and ADAPT groups. For the remaining 18 operators, we assume that entering the plateau phase occurs when a total improvement of 85% (difference between the first and the best CT) is reached for the first time. Beyond that, CTs are assumed to fall within the (upper) limit of the remaining 15% of improvement. This threshold was set based on the collected data. 77% of the CTs after the defined plateau entrance are within the expected interval. Thus, for each subject, the plateau CT is calculated as follows: (5) $C{T}_{\mathrm{plateau}}=C{T}_{\min }+0.15\cdot \left(C{T}_{\mathrm{ij}1}-C{T}_{\min }\right)$(5) where $C{T}_{\mathrm{plateau}}$ represents the max. CT for an operator to enter the plateau phase, $C{T}_{\mathrm{ij}1}$ is the first, and $C{T}_{\min }$ is the shortest (best) CT of an operator. The analysis revealed that, on average, the subjects entered the plateau phase at cycle 6.1 ± 3.2, with $C{T}_{\mathrm{plateau}}$ = 305.8 ± 86.3 sec. ADAPT group has the lowest plateau CT (263.5 ± 38.8 sec, entered at cycle 6.5 ± 2.4), followed by PROJ (303.6 ± 80.8, at 6.0 ± 3.4), PAPER (311.8 ± 113.7, at 5.8 ± 2.5), and ANIM (341.3 ± 92.0, at 6.3 ± 5.4); however, the difference between ADAPT and ANIM is non-significant (p < 0.085, d = 1.103), as in other pairwise comparisons. Thus, H2 is rejected.

4.1.3. Retention

We use two measures to assess retention in the plateau phase. The first one is the percentage of CTs below the plateau CT in Eq. (5). The second one is the memory decay exponent, indicating the level of interference, and obtained from fitting the interference-adjusted model (IAWLC) by Jaber et al. (2021), Eq. (2), to each subject’s learning data. These two measures have a strong negative correlation (−0.62) over the entire population, i.e. more CTs below the threshold indicates less interference. The percentage of CTs below the threshold is significantly lower with ADAPT (57.5 ± 22.5%) compared to ANIM (84.8 ± 15.2%; p < 0.046; d = 1.422) and PROJ (85.3 ± 10.5%; p < 0.033; d = 1.582), and non-significant compared to PAPER instructions (73.3 ± 31.3%). On the other hand, ANIM has the highest decay exponent ($\gamma$ = 0.317 ± 0.078), followed by PAPER ($\gamma$ = 0.298 ± 0.280), ADAPT ($\gamma$ = 0.253 ± 0.050), and PROJ group ($\gamma$ = 0.147 ± 0.112). The difference between ANIM and PROJ groups is significant (p < 0.024; d = 1.759). Thus, H3 is rejected.

4.2. Disability effect

Figure 4 presents the mean CT for each type of disability and cycle (Eq. 4).

Figure 4. Effect of type of disability on mean and SD (2σ) of CT in each cycle.

Illiterate has longest first cycle time. Illiterate and psychosocially disabled people have steeper initial learning curve. Cognitively disabled people learn much slower, and their cycle times show high variation. Each curve levels out until cycle 15.

4.2.1. First cycle time

The mean of the first assembly CTs is the shortest for COG (936.5 ± 227.0), followed by PSYSOC (1016.9 ± 492.6) and ILLIT group (1146.0 ± 447.7). This order is in line with H4; however, each pairwise difference is non-significant. In the first cycle, the mentor helped PSYSOC group (3.75 ± 2.605) significantly less often than ILLIT (6.75 ± 2.765; p < 0.021; d = 1.117) and COG (7.25 ± 5.092; p < 0.053; d = −0.865) groups.

4.2.2. Plateau cycle time

COG and PSYSOC groups have equal plateau CT, but the first-mentioned reached it 1.7 cycles later (293.2 ± 114.8 sec, entered at cycle 7.7 ± 2.1; and 295.4 ± 95.6, at 6.0 ± 3.4, respectively). ILLIT group has a slightly longer CT than the other two but entered the plateau earlier (319.5 ± 78.1, at 5.6 ± 3.5). The CTs (and the number of cycles needed) to enter the plateau phase were not significantly different between the types of disability, and thus H5 is rejected.

4.2.3. Retention

In the plateau phase, the percentage of CTs below the threshold is the lowest for COG (63.3 ± 28.7%), followed by PSYSOC (66.4 ± 23.5%) and ILLIT group (86.9 ± 18.8%). The percentage is significantly lower for PSYSOC (p < 0.042, d = 0.968) but non-significantly lower for COG (p < 0.069, d = 1.099), compared to ILLIT. COG disabled subjects experienced the greatest interference (γ = 0.426 ± 0.191), followed by PSYSOC (γ = 0.242 ± 0.209) and ILLIT (γ = 0.211 ± 0.112). Interference was significantly higher for COG than ILLIT (p < 0.021, d = −1.607). Based on these results, H6 can be validated.

5. Discussion

In this section, we discuss the results for each hypothesis, with the aim being to gain a better understanding of the factors behind them. We then provide managerial implications concerning the suitability of the forms of instruction for each type of disability. The discussion ends with research implications and the study limitations and future research priorities.

5.1. Hypotheses and results

H1: More advanced instructions have a positive impact on the first assembly CT. The mean of the first assembly CTs is the shortest with projected instructions, followed by animations and paper instructions.

Result: The mean of the first assembly CTs was shorter with projection than with paper instructions and significantly shorter than with animations.

This finding implies that to enhance initial learning, projecting instructions on the tabletop with blinking lights is beneficial. ANIM require additional time for understanding because of the moving images. They are mainly harder to interpret in terms of spatial awareness because of their dynamic character. When PAPER is used, the subject must scan and filter the information, which takes some time. In addition, the subject must switch attention from the instructions to the working space and vice versa, resulting in lower productivity. However, the results did not indicate a significant difference between PROJ and PAPER. Further analysis showed that those with projected instructions were significantly more dependent on mentor guidance in the first cycle than the other groups. When we look into the results in greater detail, it was observed that two outstanding subjects were part of the group that was supported with PAPER. Their first CT was by far shorter than the others’ and they did not need mentor help. As the subjects were assigned to groups randomly and there is a very high SD in this group, the results would give a significant difference when the number of subjects was increased.

H2: More advanced instructions have a positive impact on CT when learning plateaus. After a percentage (85%) improvement, the mean of the CTs at which learning plateaus is the shortest for adaptive projection, followed by projection, animations, and paper instructions.

Result: The differences between the forms of instruction were non-significant.

The results, although non-significant, mostly followed the order of the hypothesis. ADAPT shows the best and PROJ the second-best result in terms of CT at the plateau level. Filtering the amount of information helps the user to work more accurately. Only relevant information is shown, which is less intrusive and yet helpful. On the other hand, ANIM draws the user’s attention, even when the user does not need the information. The dynamic character causes an additional extraneous load. The differences in the number of cycles to entrance into the plateau phase are too low for any further conclusions to be drawn, and besides that, there is much more variation for the group that received ANIM. It might be the case that instructions in this form are not well received by every learner.

H3: In the plateau phase, instructions are no longer needed, and workers perform equally. Retention in the plateau phase in terms of the percentage of CTs below the threshold and the level of cognitive interference is the same with each form of instruction.

Result: The percentage of CTs below the threshold is significantly lower with adaptive projection compared to animations and projection. The level of cognitive interference is significantly higher with animations than with projection.

Although the use of ADAPT results in lower CT at the plateau threshold, there seems to be much higher variation in the CT after entering this phase. Considering the background of the subjects, this might indicate that they are not always confident when information is filtered. They were not able to get information on demand. Consequently, when they required additional information, they had to consult the mentor. This caused quite an increase in CT, which explains the higher deviation. As expected, cognitive interference is the highest with the form of instruction that is most intrusive.

H4: Illiteracy slows down the comprehension of new learning material. The mean of the first assembly CTs is longer for illiterate subjects than for psychosocially and cognitively disabled ones.

Result: The differences between the types of disability were non-significant.

The results, although non-significant, followed the order of the hypothesis. The inefficiency of PAPER was evident in the first cycle for ILLIT, as text would only interfere with them, while images are not enough to instruct them. Perhaps surprisingly, COG disabled were the fastest in the first cycle, with the lowest within-group variance. Further analysis showed that they and the ILLIT subjects were significantly more dependent on the mentor’s guidance than PSYSOC disabled ones. COG disabled needed guidance more frequently with more advanced instructions, PSYSOC with projection-based instructions, whereas no clear trend was observed for ILLIT. It is important to note that the introduction of the HIM unit was very brief, which led to more questions about the system and less on the job itself.

H5: Cognitive disability has a negative impact on CT when learning plateaus. After a percentage (85%) improvement, the mean of the CTs at which learning plateaus is the longest for cognitively disabled subjects, followed by psychosocially disabled and illiterate subjects.

Result: The differences between the types of disability were non-significant.

This result should be interpreted with caution as plateau analysis excluded five COG and one PSYSOC disabled subjects with insufficient numbers of cycles. With the reduced sample size, a comparison between all disability types can be only performed for PAPER and ANIM groups. A closer look at the CTs revealed that those with PSYSOC and COG issues reached an average best CT 13% lower than ILLIT with PAPER, but with ANIM it was the opposite, and their best CT was 43% higher. As Eq. (5) shows, the best CT mostly explains the plateau CT. This result indicates that ANIM are effective in training ILLIT subjects, whereas its dynamic objects disrupt the learning processes of those with PSYSOC and COG issues.

H6: Retention in the plateau phase is the worst (the percentage of CTs below the threshold is the lowest and the level of cognitive interference is the highest) for cognitively disabled subjects, followed by psychosocially disabled and illiterate subjects.

Result: The percentage of CTs below the threshold is significantly lower for psychosocially disabled compared to illiterate subjects. The level of cognitive interference was significantly higher for cognitively disabled than illiterate subjects.

These results suggest that those with COG and PSYSOC issues are less capable than ILLIT of retaining the skills acquired. A closer look at individual learners revealed that variation in performance in the plateau phase (and earlier) was especially high for the two subjects who were the best of all in the first cycle (both used PAPER). Even though they did not even need support at all in the first cycle, no mentor observations were provided to explain the unusual learning behaviour in the experiments. According to background information, one of them was a slow learner (COG) and the other was a dependent learner (PSYSOC). Their performance continued to fluctuate after the 15 cycles that were analysed, with no visible improvement. Further investigation would perhaps show whether a lack of skill or engagement in the learning task or negligible initial support would cause such unpredictable performance.

5.2. Industry 5.0 managerial implications

In this section, we provide some managerial implications of the results, in line with the Industry 5.0 concept of individualised support. Some additions to the results are given to clarify our opinion on the results. Those are mainly based on observations the mentor made during the experiment that was conducted.

5.2.1. Illiterate

The first cycle time of the ILLIT group, divided over the four forms of instruction, was the highest. Over the cycles, they almost reached the assembly time of the subjects in the other groups. This means that they have the greatest potential in terms of productivity, but they face difficulties when a new procedure must be learned. For those who are illiterate, it is very important to tailor the instructions to their needs when they learn new assembly tasks, i.e. avoid text and make the instructions as intuitive as possible. The mentor’s remarks are: ‘can’t read instructions, so had a lot of trouble with it’ and ‘went better once the procedure was known’. For this group, text-based instructions led to much worse productivity in comparison with more advanced instructions. In this study, the ILLIT group was perhaps best supported with ANIM, which includes pictorial elements that are helpful for those with impaired communication skills (Mark, Rauch, and Matt 2022). By extending PROJ with ADAPT some improvement is observed, by a lower CT when plateauing, but at a cost of worse retention thereafter. However, the differences at the plateau level are not significantly high.

5.2.2. Psychosocial disability

It is hard to draw strong conclusions about the PSYSOC disabled group. First of all, there is a great difference in cognitive abilities among them. Furthermore, these subjects depend much more than the others on their immediate feeling. However, from the results it is clear that animations are to be avoided since people with PSYSOC issues are ‘easily disturbed’ (mentor’s remark) by moving images. Both for the first and the plateau CT, ANIM provided the least productive results and PROJ the best results. As with ILLIT group, PROJ extended with ADAPT slightly reduced the plateau CT but at a cost of worse retention. A less independent operator with PAPER instructions was exceptionally fast in the first cycle but her CTs scattered to a great extent thereafter. This suggests that some people with PSYSOC disability feel especially comfortable when they are assigned instructions they are familiar with. However, to increase their ability to maintain mental capability and working pace, they often need prompts (Furniss et al. 1999), such as projections. This was also reported by the mentor when observing a subject using the ADAPT system: ‘uncertain at first, but after a few iterations it went better’.

5.2.3. Cognitive disability

People who are COG disabled learn much more slowly than others. Although their first CT was the best (on average) over the different forms of instruction, they had the smallest margin of improvement. Just like the previous group, animations are not recommended for this group. The quickly-changing images are hard to capture and translate to reality. The best performance was achieved using PAPER instructions. Not only was the CT the shortest for the first cycle, but also the subjects were more independent. Because of their disability, they experience a higher intrinsic workload, which consequently should be taken into account when instructing them. Both subjects with PAPER kept assembling more than twice as many products as their peers with projection. Some of the remarks the mentor made with this group using projection-based instructions: ‘does not work alone’; ‘has a lot of trouble understanding the instructions’ and ‘continuously struggling’ with certain components, which is in line with Korn, Schmidt, and Hörz (2013). The COG disabled persons worked very unpredictably, meaning that no clear learning effect is visible when plotting the data. Subsequently, they do not benefit from adaptivity (which is based on predictivity). The barrier to getting used to the HIM unit is also quite high for them. Because of their deficits in controlling cognition, the adaptive algorithm may also shift their experience level too soon, resulting in discomfort and a reduction of productivity.

5.2.4. General remarks

In order to support impaired assembly operators with their job, work instructions should be tailored to the individual’s needs. No form of instruction is the best; it all depends on the user (Mark, Rauch, and Matt 2022). The trend towards more advanced CAS for guiding operators has already proved its value in the regular market (Petzoldt et al. 2020); however, in the case of impaired people the use of such systems must be handled with care. During the different experimental runs, it became clear that often the intuitive and adaptive CAS forms a barrier for the people rather than a support. This is probably caused by the short introduction to the system the subjects received. Therefore, we strongly recommend providing enough time to introduce the new CAS to the operators, since this is crucial for the system to be effective.

In general, PAPER, along with ADAPT instructions, showed the best plateauing effect over the three disability groups. There might be two reasons for this. First, an operator who knows the task at hand can easily neglect PAPER instructions and is well served by the ADAPT system, where ANIM interfere with the task at hand. This is an example of an extraneous cognitive load, meaning that instructions should only be provided when it is necessary. Second, the use of PAPER is quite demanding, since the operator must search actively for the right information. This additional effort results in a higher CT at the beginning, but it also enforces retention, which, in the following iterations, helps to speed up the learning process. Challenging operators does not result in quick wins, but in the long run, it can be beneficial with respect to their productivity and independence.

Even though no serious improvement in productivity is observed using PROJ and ADAPT instructions, the mentors declared an improvement in the independence of the impaired workers. On top of that, the technology enables some of them to execute more complex tasks. This could have a great impact on the employment rate of people with a disability. To summarise the managerial implications, high-tech CAS has the potential to support the inclusion of people with various learning disabilities in industrial production. This directly contributes to the Industry 5.0 agenda on human-centric and socially sustainable production (Ivanov 2023), and the UN sustainable development goals of industry, innovation, and infrastructure (SDG 9), and reduced inequalities (SDG 10) (Bai, Zhou, and Sarkis 2023).

5.3. Industry 5.0 research implications

This study contributes to the Industry 5.0 literature (Mark et al. 2019) by practically implementing a CAS in the industry for people with various types of disabilities. The assembly learning data obtained from the experiment is valuable in the POM research, which develops and uses LCs to measure human performance in repetitive tasks (Glock et al. 2019). The data and the results have implications for developing LC theory where the form of instructions and type of disability are the explanatory factors. The results can further be used to develop the methodologies for the selection of CAS based on parameters that consider the type of disability and work task (Mark, Rauch, and Matt 2022). The findings of our study imply that high-tech CAS can have a positive impact on the employment of people with cognitive issues, in line with the review of (Sauer, Parks, and Heyn 2010).

5.4. Limitations and future research priorities

Despite its interesting findings, this paper has some methodological limitations that should be addressed in future research. First and foremost, some results are statistically non-significant and cannot be generalised. More studies with larger sample sizes would give more reliable evidence. For example, the plateau CT comparisons showed a very large effect size d = 1.103 and practical significance in favour of adaptive instructions over animations. However, to have enough statistical power (of 0.8) the group sample sizes should be increased from N = 4 to N = 11. In the field of psychology, a large effect size of d = 0.8 is assumed. This corresponds to a group sample size of N = 21, which, in our experimental setting, requires 84 participants (28 of each type of disability), which would be extremely time-consuming, if not impossible.

Another limitation is a rather short experimental time of 2 h for novices. The instruction effect should be tested over a longer training period to allow disabled workers to overcome the learning barriers of new technologies and enable gradual improvement. Adaptation to individual needs could be performed by alternating between traditional paper instructions and the most promising advanced form of instruction. In addition, instructions should be tailored to respond better to the behaviour and needs of a specific disability, such as a cognitive one.

Worker independence received rather little attention in this study despite it being an important performance indicator for disabled subjects. In further studies, mentor support (both quantity and quality) should be registered more systematically over repetitions, perhaps using automatic detection. The next limitation of our study is the lack of subjective evaluation of user experience and perceived workload. During the experiment, we learned that standard questionnaires did not work, hence, further research should develop specific methods to enable a more realistic assessment of disabled workers’ experience and perceived workload.

6. Conclusions

In this paper, we studied the effect of both the form of instruction (paper-based, animations, projection, adaptive projection) and worker disability (illiterate, psychosocial, cognitive) on the learning of a manual assembly task. The operators assembled the adjuster of a Dorel’s Maxi Cosi car seat. Times were registered and observations were made during the assembly runs. The learning effect, in terms of cycle time, was analysed in three phases: the first cycle, when entering a plateau level, and retention thereafter. Although each individual had different needs, some general conclusions can be drawn based on the results of this study:

1. More instructional elements and cognitive assistance (projection-based) enhance the first cycle.

2. Projected instructions are useful but make operators more dependent on the guidance system. If the operator’s profile allows it, challenging him/her by filtering the instruction content with increased experience leads to more independence and a better understanding of their tasks. People with psychosocial issues, on the other hand, do not feel comfortable when the level of assistance drops.

3. Highly intuitive and adaptive instructions do not seem appropriate for cognitively disabled operators with unpredictable performance and great dependence on mentor support.

4. The form of instruction should be considered individually and in an unprejudiced manner, as the differences between operators and adaptation to cognitive assistance systems vary a lot, including within the same disability group.

The experimental methods and findings contribute to the Industry 5.0 individualised human-machine interaction framework. The used cognitive assistance system with AI and spatial AR technologies enables automatic monitoring of work and projecting those instructional elements that each operator needs in each work step. Theory, experimental results, and experiences in a sheltered workplace provide insights into how to implement support for a range of disabled operators. Intuitive and adaptive instructions enable more people to become involved in complex tasks and have a better chance of being employed in the regular labour market. The advantages of cognitive assistance systems also hold true for regular workers, especially when they are constantly facing new products. The cost-efficiency they provide can be significant as companies can increase cross-training while releasing mentors for other tasks.

Acknowledgements

We thank Ton Janssen, Liese van Oort, and Rob Kerkhofs, at Senzer, and Bart Lamberigts, at Arkite NV, for providing resources and support during the project. We also thank the anonymous reviewers for their constructive comments to improve this paper.

Data availability statement

The data that supports the findings of this study is available from the corresponding author upon reasonable request.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported by the Finnish Work Environment Fund [Työsuojelurahasto] under grant 200224.

Notes on contributors

Jaakko Peltokorpi

Jaakko Peltokorpi is a Post-Doc Researcher at Aalto University, Department of Mechanical Engineering, in the Materials to Products (M2P) research group. He received a Ph.D. degree with the title ‘Worker Coordination, Collaboration and Learning in Make-to-order Assembly Production’ at Aalto University in 2018. Since then, he has been investigating and developing analytical learning curve models which also consider the cognitive load individuals encounter during assembly work. He is currently working in the ‘Green Factory – towards carbon neutral production (GREEF)’ project in the Sustainable Manufacturing Finland programme, funded by Business Finland.

Steven Hoedt

Steven Hoedt is a Master of Science in automation engineering and a Ph.D. candidate at Ghent University. The focus of his Ph.D. studies in Industrial Engineering and Operations Research is mainly on the assessment of the learning and forgetting of manual assembly operators and how to tackle the forgetting effect via training or on-the-job support systems. In addition to his Ph.D. research, he coordinates various research projects that mainly focus on human-augmented technology in the manufacturing industry and is an active researcher within the FlandersMake@UGent-ISyE research group.

Thomas Colman

Thomas Colman is a Master of Science in electronics engineering and works at the data analytics team at Kion Mobile Automation. During and for this research project he has been developing an algorithm, commissioned by Arkite NV, that monitors the learning curve of a certain test person, determines a personal experience level, and adjusts the amount of information shown accordingly. By using his experience in practical research, he played a key role in setting up, conducting, and maintaining the physical field tests.

Kim Rutten

Kim Rutten has a Master’s degree in chemical engineering from the Technical University of Eindhoven in the Netherlands. After ten years of innovation engineering, he joined Arkite, a company specialising in Augmented Reality guidance systems for the manufacturing industry. Aside from proprietary 3D detection algorithms, he focused on mining operator performance data. He is currently working as a product manager for Machine Learning at Software AG. Leveraging machine learning capabilities for leading platforms for IIoT – Cumulocity IoT, for self-service analytics – TrendMiner, for streaming analytics – Apama and the Thin Edge revolution.

El-Houssaine Aghezzaf

El-Houssaine Aghezzaf is a professor of industrial systems engineering and operations research at the Faculty of Engineering and Architecture of Ghent University. He holds a Master of Science degree and a Ph.D. in applied mathematics and operation. He currently heads the department of Industrial Systems Engineering and Product Design and has an active role in Flanders Make, a strategic research centre in Flanders. His main research interests are in integrated optimisation and simulation approaches to the design, planning, and control problems arising in manufacturing systems and in logistical and utility networks.

Johannes Cottyn

Johannes Cottyn is an assistant professor in industrial automation at Ghent University in the department of Industrial Systems Engineering (ISyE) and Product Design. He leads the FlandersMake@UGent-ISyE research group within the Flexible Assembly of Flanders Make research cluster. His initial research interests lie in the combination of industrial control and software systems (e.g. MES) and manufacturing excellence best practices (e.g. Lean). His current research activities are focused on human-augmented and digital-twin-enabled manufacturing.

Notes

1 HIM installed at Senzer (sheltered workplace in NL). Retrieved 21 February 2023, from https://www.youtube.com/watch?v=YF9Fbvejp-s.

2 Arkite Platform. Retrieved 21 February 2023, from https://www.arkite.com/platform/.

Annex A

Principles, their explanations, and related guidelines to handle various types of cognitive load (Clark, Nguyen, and Sweller 2011; Li et al. 2018).

Table

Principle	Explanation	Guideline
Avoiding extraneous load
Worked examples	A demonstration that illustrates the steps to complete the task.	Use some steps as worked examples and others as practice problems and gradually move on to mere practice problems.
Focus attention	Directing cognitive resources toward relevant content in the instructions.	Use cues (e.g. arrows) and signals (e.g. bolding) to focus attention on relevant visual and textual content.
Split attention	Cognitive load resulting from integrating two or more dependent sources of visual information that are physically separated.	Integrate explanatory text and related visuals close to each other.
Modality	Working memory has a channel for both auditory and visual information. With dual modality, performance on the primary task is not significantly depressed.	Present text in auditory rather than written mode when integrated with other visual material, such as figures, animations, or video.
Redundancy	Identical information formats that do not contribute to schema acquisition or automation interfere with learning. Less is more.	Break information down to essentials. Eliminate extraneous text, visuals, and audio. Eliminate redundancy in information delivery modes.
Performance aids	Many tasks do not require accessing knowledge from memory but can use various reference resources to supplement internally stored knowledge and skills.	Use graphics alone when the task can be effectively communicated visually. Use arrows or other motion cues rather than text to depict motion. Present graphics from the visual perspective of the learner.
Fading of memory support	Fade the support as learners gain expertise.	Gradually remove memory support (instructions) as training progresses.
Managing intrinsic load
Pre-training	Allowing the learner to build schemata or prior knowledge about the process (concept and components of a system).	Segment a system into components for meaningful learning. Visually segregate and name each component. Represent their state change in the context of the whole system.
	Allowing the learner to build schemata or prior knowledge about procedure (steps to complete the actual task).	Two strategies for steps: (1) teaching, practising, and again teaching with supportive information; (2) teaching supportive information and then teaching each step hands-on.
Learners control over pacing	Giving learners the opportunity to assimilate new content at their own pace.	Implement self-pacing in instructions with ‘continue’ or ‘next’ buttons. If not possible, try to develop instructionally paced environments.
Promoting germane load
Self-explanation	A mental dialogue at a worked example that helps learners to understand the example and build a schema from it.	Get learners to engage in good self-explanation while they read instructions.
Automaticity	The built schemas need to be automated so that cognitive resources are freed to allow the construction of more complex schemas.	Help learners to automate skills by giving them many opportunities to rehearse new knowledge and skills, through hands-on practice or mental rehearsal of the steps.