Review of task allocation for human-robot collaboration in assembly

ABSTRACT

Due to the high cost pressure and the increasing variant diversity, the cooperation of humans and robots represents a promising technological solution to achieve higher flexibility and efficiency in assembly. It is therefore attracting significant interest from both researchers and practitioners. As a result, numerous reviews have been published addressing different aspects of human-robot collaboration, such as safety, interaction modalities, programming, and applications. However, in this paper, for the first time, the aspect of task allocation for collaborative assembly is methodologically examined through a systematic literature review. This paper presents the current state of the art in task allocation approaches, investigates the criteria for deciding on a suitable task assignment, and discusses challenges and future research needs. After filtering the 521 publications that resulted from the initial search process, 37 relevant publications were included in the analysis and grouped into a proposed classification consisting of two main categories, static and dynamic task allocation approaches. Based on the results of the literature review, this paper presents a reference model for human-robot collaborative assembly.

KEYWORDS:

• Assembly
• human-robot collaboration
• task allocation
• process planning
• industry 4.0
• smart manufacturing

1. Introduction

Industrial assembly must cope with a high degree of complexity, as the configuration of product variants mostly takes place in assembly (ElMaraghy et al. 2013). The underlying external market drivers, such as stronger customer orientation and individualization of products, lead to increasing product varieties and decreasing lot sizes. This strongly increases the complexity of industrial assembly as more variants with fewer equal parts are assembled on the same lines (Bänziger, Kunz, and Wegener 2018). In addition, the available amount of well-qualified employees on the market is decreasing due to demographic changes (Bänziger, Kunz, and Wegener 2018), while changing market demands require scalability of production volumes. Furthermore, the increasing cost pressure in industrial assembly (Lotter and Wiendahl 2012) has dramatically restricted the profit margin for companies, mainly because of increasingly expensive human labor costs (Chen et al. 2014). Those trends require both the use of automation technology (Müller, Vette, and Mailahn 2016) and a flexible assembly (Scholz-Reiter and Freitag 2007). While other manufacturing systems in industry are already widely automated, assembly processes remain predominantly operated manually (Spena et al. 2016; Fast-Berglund et al. 2016), especially for low to medium product volumes (Lotter and Wiendahl 2012). In conjunction with the aforementioned flexibility demands, less defined working procedures in low-volume assembly compared to fully automated robotic cells (Antonelli and Bruno 2019) result in traditional automation technologies with limited flexibility to be unsuited (Ranz, Hummel, and Sihn 2017).

To meet those challenges, Industry 4.0 offers new technologies such as assistance systems and automation solutions to improve productivity, quality, ergonomics, and the flexibility of manual and hybrid assembly processes. These technologies cognitively and physically support assembly workers (Mark, Rauch, and Matt 2021). Cognitive support for workers in manual assembly is particularly enabled by informational assistance systems that guide assembly workers through the process, ensuring high quality and enabling less-qualified workers to perform the assembly (Hinrichsen et al. 2016). Regarding physical support of assembly processes, collaborative robots belong to the key enabling technologies within Industry 4.0 (Antonelli and Bruno 2019). Unlike traditional industrial robots, which operate behind separating safety barriers, collaborative lightweight robots (Colgate, Wannasuphoprasit, and Peshkin 1996; Krüger, Lien, and Verl 2009) allow workers and robots to work closely together in hybrid workplaces. This expands the applicability range of robotic automation towards lower production volumes (see Figure 1) (Kopp, Schäfer, and Kinkel 2020; Matthias and Ding 2015).

Figure 1. Field of application for human-robot collaboration in assembly (in accordance with and adapted from (Lotter and Wiendahl 2012)).

The introduction of collaborative robotic systems into manual assembly stations offers the potential to combine the advantageous skills of human workers, such as dexterity, flexibility, perception, and intelligence, with those of robots, such as precision, repeatability, and strength (Tsarouchi, Makris, and Chryssolouris 2016; Fast-Berglund et al. 2016). Especially in repetitive tasks, robots are not affected by performance fluctuations and fatigue, unlike humans (Bauer et al. 2016). Moreover, the integration of collaborative robots into manual workstations decreases the physical workload of human workers and allows the worker to focus on tasks that require human-specific capabilities. Additionally, the risk of workplace injuries such as musculoskeletal disorders can be reduced. Workplace injuries are one of the major reasons for absenteeism and productivity loss in the manufacturing industry and are often caused by frequent repetitions of unhealthy postures over a long period of time (Makrini et al. 2019). Therefore, the vision of human-robot collaborative assembly is to increase quality, efficiency, flexibility, and productivity, while also improving ergonomic working conditions and reducing repetitive workload (c.f. (Müller, Vette, and Mailahn 2016; Bughin et al. 2018). From an economic perspective, various studies show that, if appropriately designed, human-robot collaborative assembly is beneficial for variable productions with frequent model and volume changes (c.f. (Takata and Hirano 2011)) and reduces overall production cycle times (Antonelli, Astanin, and Bruno 2017) as well as unit costs, while increasing flexibility (Krüger, Lien, and Verl 2009).

However, the degree of automation in manual assembly is still low. Even in the automotive industry, where the use of robots is well established, it is only at an estimated 20% (Steil and Maier 2020). One explanation for this discrepancy between the expected potentials and the dissemination in practice are several open challenges, e.g. the changed way of working when introducing collaborative robots in industrial manual assembly (Bruno and Antonelli 2018; Schmidbauer et al. 2020; Qian et al. 2020). Thus, it is necessary to rethink both assembly workplaces as well as organizational structures and processes (Tan et al. 2009). The challenges to be considered for the introduction of human-robot collaboration (HRC) can be classified into safety-related, planning and design, and human factors (c.f. (Villani et al. 2018)). The former two aspects, safety assessment and planning, are supported by the industrial study from (Gaede et al. 2018), based on interviews with robot manufacturers and system integrators. Regarding the safety assessment, the required application-individual risk assessment, the application of relevant safety norms in practice, and the costly as well as time-consuming approval and certification through inspecting authorities (see also (Schmidbauer et al. 2020)) were stated as key barriers. Among the planning aspects, the identification of HRC-suitable workstations, the workplace design, the determination of suitable process plans, and an efficient task allocation between humans and robots (Bruno and Antonelli 2018; Villani et al. 2018; Wang et al. 2019, 2020) were considered as the top challenges for implementing HRC (Gaede et al. 2018; Ranz, Hummel, and Sihn 2017). Also, a lack of robot programmers in small and medium-sized companies (Schmidbauer et al. 2020) as well as a lack of well-proven methods for the correct assessment of economic profitability (Bruno and Antonelli 2018) were argued in literature. Finally, human-centered technologies and human factors such as cognitive workload (e.g. by psychological assessment of operators’ well-being, acceptance, and human-centric design), collaboration fluency (e.g. by user-friendly programming approaches or natural, intuitive information exchange and communication methods), as well as trust (e.g. by well-proven, conducive safety mechanisms) complement the set of challenges (Darvish et al. 2018; Villani et al. 2018; Baltrusch et al. 2021).

Faced with those challenges of HRC, both research and industry have developed and proposed various approaches, such as intuitive and fast robot programming environments to enable robot programming for non-experts (Niermann et al. 2022), methods to improve safety or simplify and standardize safety assessments (Marvel, Falco, and Marstio 2015), procedure models to methodologically support planners when introducing collaborative robots (Peifer and Weber 2020; Petzoldt et al. 2021), methods to support assessment of economic profitability (Rally et al. 2020), technologies for intuitive bidirectional communication between humans and robots (Gustavsson et al. 2018; Mahadevan et al. 2021), psychological aspects in the context of a collaborative assembly, e.g. effects on human stress (Ender 2021; Wiese et al. 2022), and approaches for task allocation in HRC (see chapter 4 section 4). While many review publications on HRC and several dedicated survey publications address some of the mentioned issues (see chapter 2 for detailed analysis section 2), the proposed approaches for task allocation between humans and robots in hybrid assembly have not yet been reviewed.

The problem of task allocation in HRC mainly arises because both resources (humans and robots) obtain different skill sets. In manual assembly cells, human workers are largely equally suited for performing assembly tasks. Thus, their individual skills do not need to be considered, which reduces the allocation problem to workload balancing. However, in HRC, both resources have different strengths, capabilities, and limitations regarding the execution of different tasks that should be exploited as much as possible. Moreover, there is no need for workload balancing between humans and robots because robots are not affected by fatigue effects and are able to cope with a higher workload compared to humans (Antonelli and Bruno 2019, 2018). Instead, the individual skill sets should be taken into account for creating suitable process plans and efficient task allocation to humans or robots while considering and optimizing the overall objectives such as cycle times, costs, and quality (Bänziger, Kunz, and Wegener 2018; Zanchettin et al. 2019). Despite the fundamental importance of task allocation in enabling reasonable HRC and its potential to increase productivity (S. Kumar, Savur, and Sahin 2021), it remains an open research topic (c.f. (Beumelburg 2005; Sipsas, Nikolakis, and Makris 2021)). To approach this topic and related questions, this paper reviews existing approaches to human-robot task allocation in industrial assembly to provide an extensive overview of the topic for both research and industrial practice. More precisely, this review paper answers the following two research questions:

• RQ1: What is the current state of the art of task allocation in human-robot collaborative assembly, and how can the existing approaches be classified?

• RQ2: What criteria are used to decide on the allocation of assembly tasks between humans and robots?

The remainder of this paper is structured as follows: section 2 analyses previously published, related reviews in the field of human-robot interaction as well as assembly sequence planning, and highlights identified needs for further investigation. Section 3 explains the applied methodology for conducting this literature review. In section 4, the review results are presented by firstly providing a descriptive analysis of the identified publications, secondly categorizing and explaining the individual approaches and their respective task allocation procedures, and thirdly identifying the criteria applied for task allocation. Subsequently, in section 5, the different approaches are consolidated into a reference model and the results, the identified research gaps as well as future research directions are critically discussed. Finally, section 6 concludes this research with a summary.

2. Related work: overview of existing literature reviews on human-robot collaboration

In conjunction with the high expectations placed on HRC, it is a field of research that has gained much interest during the past decade, as reflected in the large number of published articles and reviews that address various aspects of the topic (c.f. (Hentout et al. 2019; Tsarouchi, Makris, and Chryssolouris 2016)). A total of 22 review and overview papers were identified in the related work, focusing on safety (10), human-robot interaction and communication technologies (8), application domains for HRC (3), robot programming and motion planning (5), metrics and measures for evaluating HRC (2), human factors (3), as well as task planning and allocation (1) (some reviews take into account several aspects). All review publications were recently published between 2016 and 2022, and are summarized in Table 1 with respect to the reviewed aspect, distinguishing between focal topics of the respective review, indicated by ‘x’, and only marginally or briefly considered aspects marked by ‘(x)’. Here, two publications by the same authors with largely identical review content were aggregated in two cases. Further, it is worth mentioning that only eight out of the 22 identified publications conducted a systematic, thus reproducible, literature review (SLR). The following discussion of previous literature reviews on HRC is organized according to the order shown in Table 1.

Table 1. Overview of considered aspects in existing literature reviews on human-robot collaboration.

Reference (HRC-review paper)	Systematic literature review	Safety	Communication & interaction	Programming & motion planning	Applications	Human factors	Metrics & measures	Task planning & allocation
(Arents et al. 2021)	✓	x	x		(x)
(Baltrusch et al. 2021)	✓					x
(Dianatfar, Latokartano, and Lanz 2020)			x		(x)
(Zoltan and Dhir 2020)	✓	x		x		x		(x)
(Galin and Meshcheryakov 2019)		x
(Gualtieri, Rauch, and Vidoni 2021)	✓	x	(x)	(x)	(x)	x	(x)	(x)
(Halme et al. 2018)		x	(x)
(Hentout et al. 2019)	✓	(x)	(x)	(x)		(x)		(x)
(Hjorth and Chrysostomou 2022)	✓	x	x					(x)
(Inkulu et al. 2021)			x		x
(S. Kumar, Savur, and Sahin 2021)		x	(x)				(x)
(Marvel et al. 2020)							x
(Eloise et al. 2019)	✓	(x)			x
(Mukherjee et al. 2022)		(x)	x	x
(Nelles, Th Kwee-Meier, and Mertens 2019)	✓						x
(Robla-Gomez et al. 2017)		x
(Tsarouchi, Makris, and Chryssolouris 2016)			x	x			(x)	x
(Villani et al. 2018, 2018)		x	x	x	x	(x)
(Wang et al. 2019, 2020)		(x)	x	x		(x)		(x)
(Zacharaki et al. 2020)		x
Total number of considering review papers	8	14	12	7	6	6	5	6
Total number of reviews considering aspect as focus-topic		10	8	5	3	3	2	1

x: considered in review as focus-topic | (x): only marginally considered, not focus of review | ‘ ‘: not considered

As follows from Table 1, most review publications consider safety aspects, interaction modalities, and communication technologies. Similarly, (Gualtieri, Rauch, and Vidoni 2021) identified safety as the most developed research category. While most publications specifically focus on one or two of the above-mentioned subjects, a few comprehensive extensive reviews, e.g. (Hentout et al. 2019; Wang et al. 2019, 2020; Zoltan and Dhir 2020), provide a good overview of the entire field of HRC and thus represent a good starting point into the topic. However, due to their wide review scope, those reviews are less detailed within the individual aspects compared to dedicated reviews. Worth highlighting is the work of (Hentout et al. 2019), which conducted a systematic literature review within the time period from 2008 to 2017 and considered a total of 300 articles within the field of HRC, clustered into 39 categories and sub-categories, covering all essential technological areas. Another interesting work was conducted by (Gualtieri, Rauch, and Vidoni 2021), which analyzes in a review of 55 papers within the years 2015 to 2018 the importance and influence of various technological sub-clusters for and on the aspects of safety, human factors, and ergonomics. In their review, they identified a vagueness and uncertainty in the research community regarding the definition of the terms safety and ergonomics (Gualtieri, Rauch, and Vidoni 2021).

Fundamental aspects and requirements regarding safety in HRC are covered by various of the identified safety-related review publications (Villani et al. 2018, 2018; Robla-Gomez et al. 2017; Galin and Meshcheryakov 2019 S. Kumar, Savur, and Sahin 2021). On the one hand, these publications present overviews and classifications of the general safety standards for HRC, which are grouped into A- (general aspects, risk assessment, and risk reduction), B- (functional safety for different technologies), and C-type (safety requirements and modes for safe human-robot interaction) standards. On the other hand, the publications provide a detailed description of the collaborative operation modes according to (DIN EN ISO 2011, 2011; ISO/TS 2016). Accordingly, the collaborative operation modes range from safety-related monitored stop to hand-guidance, speed and separation monitoring, and power and force limiting. Based on nine relevant articles from 2009 to 2020, (Hjorth and Chrysostomou 2022) perform an analysis of the respectively applied safety mechanisms and modes for the use-case of non-destructive disassembly. Here, speed and separation monitoring is identified as the most commonly applied approach. Research aspects for achieving safety objectives were classified into collision control methods (pre- and post-collision), motion planning measures, and action prediction algorithms (S. Kumar, Savur, and Sahin 2021). Further, methods for risk measurement and assessment are discussed in (Villani et al. 2018, 2018; Zacharaki et al. 2020). (Halme et al. 2018) focus on vision-based safety systems for HRC, aiming to analyze the readiness level of vision-based safety systems for industrial requirements by assessing the use-case flexibility, systems speed, and level of collaboration of 25 papers in total. (Gualtieri, Rauch, and Vidoni 2021) clustered the 35 identified articles that aim to improve safety into methods for contact avoidance and approaches for contact detection and mitigation. Afterward, they analyzed the importance level of research themes such as motion planning, simulation, assistance systems, safety management, sensor systems for object tracking, artificial intelligence, and robot system design for these clusters. They found a higher interest in prevention (methods that operate prior to potential contact) than in protection (methods to minimize effects of occurred physical contact) within safety-related research publications (Gualtieri, Rauch, and Vidoni 2021). Moreover, the use of machine learning methods to improve safety, e.g. by predicting collisions, was investigated by (Mukherjee et al. 2022). Interestingly, even though safety is known as a crucial aspect in HRC, (Arents et al. 2021) found in their systematic literature review of 46 papers that a quarter of all studies did not apply any of the above safety measures for cooperative operation. More than half did not even apply the safety standards when human and robot were collaboratively working on the same object.

Regarding the modalities for human-robot interaction and applied technologies for communication, one can distinguish between one-way and two-way (bidirectional) communication between humans and robots (Inkulu et al. 2021). Inkulu et al. (2021) suggest that, depending on the proximity of the collaboration, different modalities and technologies for communication should be applied. User interfaces can be clustered into visual, auditory, and haptic communication, with the former including communication methods in virtual, augmented, and mixed realities (c.f. (Tsarouchi, Makris, and Chryssolouris 2016; Villani et al. 2018, 2018; Hjorth and Chrysostomou 2022 S. Kumar, Savur, and Sahin 2021)). Human-to-robot communication methods can be classified into explicit (i.e. the human commands the robot) and implicit (i.e. the robot interprets the human state) approaches (S. Kumar, Savur, and Sahin 2021). The advantages and disadvantages of the different modes of input were analyzed by (Mukherjee et al. 2022). (Arents et al. 2021; Inkulu et al. 2021) analyzed different kinds of sensors and devices used for human-to-robot and robot-to-human communication and concluded that 3D cameras (c.f. (Halme et al. 2018)) and force/tactile sensors are the most commonly used sensors (Arents et al. 2021). Also, the information monitored using these sensors is examined. To collect information about the human worker during task execution, algorithms for human detection and tracking, gesture recognition, speech recognition, human physiology recognition, and facial recognition as well as force detection methods on robot side are applied and used as an input for motion planning and collision avoidance algorithms. Additionally, (Dianatfar, Latokartano, and Lanz 2020) investigated the use of virtual (VR) and augmented reality (AR) in HRC by assessing 19 articles that propose such solutions, and analyzing the tools and hardware devices to enable VR/AR.

The topics of robot programming and motion planning for HRC are of fundamental importance for enabling the usage of robots in collaborative scenarios. Due to the presence of the human operator, motion planning needs to adjust or react accordingly to human movements. The central role of motion planning and control, along with safety and communication, is pointed out as an essential prerequisite for HRC by the review of (Gualtieri, Rauch, and Vidoni 2021). They found that most of the reviewed articles (25 out of 55 articles) are related to motion planning and control. In addition, they analyzed the importance of motion planning and control for safety-related aspects (contact avoidance, contact detection, and mitigation) and for human factors (physical ergonomics, cognitive and organizational ergonomics) and found that motion planning and control was the only topic of importance for all aspects. Furthermore, robot programming approaches, architectures, and frameworks in the context of HRC have been reviewed by (Zoltan and Dhir 2020; Villani et al. 2018, 2018; Tsarouchi, Makris, and Chryssolouris 2016) with various proposed classifications of programming approaches (c.f. (Petzoldt et al. 2021)). Moreover, a dedicated review on robot learning strategies for HRC was conducted by (Mukherjee et al. 2022), covering supervised, unsupervised, and reinforcement learning techniques and focusing especially on learning by demonstration.

Application scenarios of HRC are marginally considered in some of the mentioned reviews, e.g. (Arents et al. 2021; Dianatfar, Latokartano, and Lanz 2020; Gualtieri, Rauch, and Vidoni 2021) – typically by mentioning the use case for the individual approaches in a table. However, since the publications are screened and filtered according to relevant criteria for the considered specific topic of the review, they only cover a partial subset of the HRC applications. In contrast, three review publications (Eloise et al. 2019; Inkulu et al. 2021; Villani et al. 2018) directly address application scenarios for HRC. The industrial use-cases of handling, welding, assembly, and automotive (Villani et al. 2018), or assembly, human assistant, quality control, and machine tending (Eloise et al. 2019), respectively welding, assembly, inspection, and other (material handling, palletizing, painting) (Inkulu et al. 2021) were found as application fields for HRC in industry; with assembly being the most studied task (Eloise et al. 2019). Moreover, based on a review of 41 articles (35 unique use cases) from 2009 to 2018 (Eloise et al. 2019) identified the automotive industry (23% of studies) and the electronics industry (17% of studies) as the major industrial sectors.

The consideration of human factors and ergonomics in HRC has been growing significantly in recent years (c.f. (Gualtieri, Rauch, and Vidoni 2021)). This is also reflected by the number of review papers related to human factors: While (Villani et al. 2018) only shortly mentioned the importance of taking into account human factors, such as mental stress and anxiety induced by close interaction with robots in their survey paper from 2018, recent systematic literature reviews on the topic of human factors in HRC were conducted and published in 2020 and 2021 (Baltrusch et al. 2021; Zoltan and Dhir 2020; Gualtieri, Rauch, and Vidoni 2021). By analyzing nine exploratory studies and reviews on this subject, (Baltrusch et al. 2021) dedicatedly examined the effects of HRC on aspects of job quality. They identified four relevant categories encompassing both social and psychophysical aspects, i.e. cognitive workload, collaboration fluency, trust, as well as acceptance and satisfaction. Regarding collaboration fluency, (Baltrusch et al. 2021) highlight the importance of adapting to the human. Similar aspects of human factors were also pointed out by (Zoltan and Dhir 2020), who identified in their general review of 87 articles on HRC in production a total of 16 papers covering human-centered subjects. Those were categorized into eight categories: adaptation, leadership, trust, workload, attitudes, frameworks, exchange force, and holding force (Zoltan and Dhir 2020). (Gualtieri, Rauch, and Vidoni 2021) categorize human-centered aspects into physical, cognitive, and organizational factors and conclude in their review that related topics to human factors are considered emerging but not yet mature (Gualtieri, Rauch, and Vidoni 2021). They also emphasized the need and importance of respective developments for cognitive and organizational support, aiming to increase trust and acceptance, which represent a major factor for the efficiency of a collaborative human-robot production system (Gualtieri, Rauch, and Vidoni 2021). Even though (Arents et al. 2021) did not focus on human factors in their review, they highlighted the importance of both proper safety systems and well-established bidirectional communication methods for achieving trust. This requires explainable, predictable, and understandable robot actions (Arents et al. 2021).

For assessing the performance of a collaborative workstation, metrics and measures are required. The review on evaluation metrics from (Nelles, Th Kwee-Meier, and Mertens 2019) identified a total of 30 publications addressing this subject and proposes a clustering of the identified metrics into those for measuring human well-being and human-robot system performance. They divide methods for measuring those metrics into questionnaire-based methods and psychophysical measurement methods (Nelles, Th Kwee-Meier, and Mertens 2019). (Marvel et al. 2020) present an even more comprehensive collection of metrics and measures for the evaluation of human-machine interfaces and human-robot interaction. They suggest sets of quantitative (33 different metrics) and qualitative (21) metrics for evaluating HRC. These metrics consider the assessment of characteristics related to either the team, the operator, the robot, or the process. Moreover, a list of objective measures, including 19 metrics to evaluate the three criteria of safety, performance, and productivity of human-robot collaborative processes, is specified in (S. Kumar, Savur, and Sahin 2021).

Finally, some reviews that consider the aspect of task sequence planning and task allocation could be identified. However, most of those reviews only marginally consider task allocation and focus on one of the topics mentioned above (Hjorth and Chrysostomou 2022; Gualtieri, Rauch, and Vidoni 2021) or provide a general overview of the research field of HRC (Hentout et al. 2019; Wang et al. 2019, 2020; Zoltan and Dhir 2020). Thereof, (Zoltan and Dhir 2020) shortly mentioned five approaches for task allocation, (Hentout et al. 2019) identified and discussed twelve papers considering task allocation and scheduling without giving more profound insights into the methodological approaches or decision logics for task allocation. (Hjorth and Chrysostomou 2022) only binary checked if task sequence planning is included in the analyzed articles on disassembly. (Gualtieri, Rauch, and Vidoni 2021) identified nine publications on task scheduling strategies but only analyzed the aspect’s overall importance for safety and human factors without describing the approaches themselves. Only one review explicitly addresses task planning, namely the work of (Tsarouchi, Makris, and Chryssolouris 2016), which considers publications up to 2015 on the topics of task planning and robot programming. However, this work has a stronger focus on robot programming approaches than on task allocation and only considers publications from the years 2005 to 2015. Furthermore, (Tsarouchi, Makris, and Chryssolouris 2016) did not perform a systematic literature review.

Apart from HRC, various reviews on assembly sequence planning, scheduling, or line balancing have been published for fully manual or fully automated assembly, e.g. (Soumen and Gagnon 1989; Rashid, Hutabarat, and Tiwari 2012; N. Kumar and Mahto 2013; Lihui et al. 2009). Also, multi-robot task allocation is a well-researched topic with various reviews and taxonomies being proposed and published in the last decades, e.g. (Gerkey and Matarić 2004; Khamis, Hussein, and Elmogy 2015; Seenu et al. 2020). However, these reviews, as well as the discussed approaches therein, do not consider HRC and are consequently not applicable for human-robot collaborative assembly due to the different capabilities and characteristics of the two resources and the limited predictability of human behavior, resulting in greater complexity in the workplace (c.f. (Bänziger, Kunz, and Wegener 2018)).

In summary, all related reviews either consider approaches for task allocation only in a short section, identify only a limited number of related papers, do not consider recent research on task allocation due to their earlier publication date, or just consider manual or automated assembly without taking into account the collaboration between humans and robots. Hence, a systematic literature review on task allocation approaches for human-robot collaborative assembly is a research gap requiring further investigation, which is addressed in the systematic review conducted in the paper at hand.

Thus, the objective and contribution of this paper is to provide a comprehensive overview of task allocation approaches for human-robot collaborative assembly through a systematic literature review. The identified approaches are categorized and analyzed in terms of the methods and allocation logics applied, as well as the criteria used as a basis for decision-making.

3. Review methodology and literature selection strategy

The systematic literature review (SLR) methodology used in this work is structured by following the approach from (Vom Brocke et al. 2009). This method was chosen because of its well-defined and recipe-alike reviewing structure that covers all important phases within a literature review instead of mainly focusing on analysis and synthesis steps (Osterrieder, Budde, and Friedli 2020). Therefore, it is frequently used in the research community for conducting SLR and has recently been applied for various literature reviews on topics related to Industry 4.0, e.g. smart factories (Osterrieder, Budde, and Friedli 2020), digital twins (van der Valk et al. 2020), or cyber-physical production systems (Xuan, Goepp, and Siadat 2020). The literature reviewing methodology by (Vom Brocke et al. 2009) comprises five main phases – (I) definition of review scope, (II) conceptualization of topic, (III) literature search, (IV) literature analysis and synthesis, and (V) research agenda –, which ensure reproducibility of this review. The five phases are detailed in the following.

3.1. Definition of review scope

In the first phase, the scope and flavor of this literature review are defined in accordance with (Vom Brocke et al. 2009). According to the characteristics and categories presented in the taxonomy of literature reviews proposed by (Cooper 1988), the six components (1) focus, (2) goal, (3) organization, (4) perspective, (5) audience, and (6) coverage are defined. Regarding the focus (1), the research outcomes and their practical application are considered, with the goal (2) of summarizing the findings in a categorized way. The procedure in this review is methodological, but by grouping and categorizing publications that propose similar approaches, this review is more conceptual regarding its organization type, as categories are proposed, existing types of approaches are discussed, and a reference model is derived in section 5 (3). To provide an objective and unbiased review, the findings are represented in a neutral manner (4). As mentioned earlier, this review addresses both researchers and industrial practitioners interested in efficient HRC in assembly (5). Finally, a complete coverage of the relevant literature (6) is provided, although certain limitations implicitly arise from the selection of literature databases and languages considered.

3.2. Conceptualization of topic

The second phase comprises the conceptualization of the topic (Vom Brocke et al. 2009) with the target of narrowing down the considered subject and identifying potential keywords that can be used as search terms and expressions for the literature search conducted in phase (III). Referring to HRC in general, the respective related reviews were analysed to ensure a broad consideration of the topic in the first place, as presented in section 2, and relevant key expressions as well as synonymously used terms were identified. For instance, the terms ‘robot’ and ‘machine’, ‘collaboration’ and ‘cooperation’, ‘task’ and ‘process’, as well as ‘planning’, ‘allocation’, ‘assignment’, ‘distribution’, and ‘scheduling’ are not used consistently in the literature. Hence, those were defined as synonymous search terms to ensure a broad coverage of literature.

3.3. Literature search

Following (Vom Brocke et al. 2009), the literature search process is conducted in the third phase. This includes four steps: (1) the selection of relevant journals covering the identified research subject and, based on this, (2) the selection of appropriate databases containing these journals. Then, (3) defining search strings and conducting the keyword-based literature search in the identified databases. In addition, (4) conducting a forward and/or backward search, as shown in Figure 2. Moreover, only peer-reviewed publications (basic quality criteria) (Rowley and Slack 2004) from relevant journals on the topics of smart manufacturing and human-machine interaction, as well as proceedings of renowned conferences of this community, are considered as sources for the search process. Based on the relevant journals, the databases for the keyword-based literature search were selected. Here, the possibility of an advanced search using Boolean operators to narrow down the search was defined as a requirement, as recommended by (Rowley and Slack 2004). Therefore, Scopus was chosen as a search database because of its broad coverage and access to the identified relevant journals from various publishers (e.g. Elsevier, Springer, Taylor & Francis, Emerald, MDPI, IEEE) as well as well-known conferences (e.g. CIRP). In addition, IEEE Xplore is used to ensure full coverage of IEEE conferences, identified in phase II.

Figure 2. Literature search process (in accordance with (Vom Brocke et al. 2009)).

Next, search strings were defined that are composed of the search terms shown in Table 2. Those were entered in both databases in the respective syntax with the search limited to title, abstract, and keywords, as given in Table 2. The database search was conducted on 27^th January 2022, and its results were exported to an MS Excel document. With the first and second search terms, the overarching topic area is delineated and specified to the investigation framework of the cooperation between humans and robots. The third and fourth search terms limit the research field to the main investigation aspect of task allocation. Modified spellings of the search terms are taken into account by using wildcards (*). Regarding the use case of (industrial) assembly, the articles were narrowed down as part of the screening process (phase IV) to avoid premature exclusion of articles through the selected key terms.

Table 2. Search terms and query strings utilized for keyword-based literature search in databases.

Search terms
1	(human* or man*) AND (machine* or robot*)
2	cooperat* OR collaborati* OR interacti* OR workplace*
3	task* OR management* OR process*
4	planning OR allocation OR assignment OR classification OR dynamic OR distribution
Full search string
Scopus	TITLE ((human* OR man*) AND (machine* OR robot*)) AND TITLE (cooperati* OR collaborati* OR interacti* OR workplace) AND TITLE (task* OR management* OR process*) AND TITLE-ABS-KEY (planning OR allocation OR assignment OR classification OR dynamic OR distribution)
IEEE Xplore	(“Document Title”:human* OR man AND “Document Title”:machine* OR robot) AND (“Document Title”:cooperati* OR “Document Title”:collaborati* OR “Document Title”:interacti*OR “Document Title”:workplace) AND (“Document Title”:task* OR “Document Title”: management OR “Document Title”:process*) AND (“Abstract”:planning OR “Abstract”:allocation OR “Abstract”:assignment OR “Abstract”:Classification OR “Abstract”:dynamic OR “Abstract”:distribution)

After reviewing and selecting the full-text articles, a forward (literature citing the articles) and backward (literature cited in the articles) search (c.f. (Webster and Watson 2002)) was performed for publications classified as highly relevant, which then complemented the considered literature.

3.4. Literature analysis and synthesis

A four-step filtering process with the articles found from the initial search query was conducted. Initially, all duplicates were removed, followed by three content-related screening rounds. Publications for the period up to early 2022 were included due to the date of the database search. For each screening round, articles were either classified as out-of-scope or assigned a relevance (low, medium, or high). Out-of-scope articles were excluded in each case, while the relevance rating was updated in each filtering round. In the first round, the screening was based on title and abstract, aiming to exclude out-of-scope records, which was done subjectively, relying on the authors’ expertise. Next, overviews, reviews, and commentaries were excluded, and only articles written in English or German were included due to the authors’ language skills. Further, only references published in journals or conference proceedings with a peer-review process were considered to meet the quality criterion for articles included in this review. In the second screening round, the full-texts of the publications were accessed, inaccessible articles were excluded, and the eligibility was assessed by skimming the respective articles. Here, for instance, articles that do not consider assembly, e.g. (Fiore, Clodic, and Alami 2016), who investigate the task of cleaning a set of furniture, or records that only mention the task allocation topic in passing, were excluded. Nevertheless, as some of these articles present relevant aspects for the subject of task allocation and human-robot collaboration in a wider context, these are still briefly analyzed and discussed in section 5.3, especially with regard to potentials for combination and improvement of current solutions as well as known issues and open challenges. In the third filtering round, a more in-depth analysis was conducted by reading the articles in detail to assess the final relevance based on the preliminary relevance determined in the previous stages. For instance, articles with low or medium relevance represent either early publications from the same authors of already included approaches (i.e. similar, less detailed articles) or approaches in which the human fully needs to decide on the task allocation, e.g. by pressing interaction control buttons that trigger either human or robot task execution as in (Maurtua et al. 2016). Such approaches are excluded as the scope of this review paper is on methodological, systematic methods, and decision logics for task allocation. Therefore, in this final filtering stage, only records marked with high relevance and relevant references from the forward and backward search (see above, phase III section 3.3) were included in the set for the qualitative literature analysis.

Following the idea of a concept matrix by (Webster and Watson 2002 Vom Brocke et al. 2009), the resulting final set of publications was analyzed by categorizing them into two main categories – static and dynamic task allocation approaches (in accordance with the distinction into offline or online allocation proposed in (Darvish et al. 2018)) – and four sub-categories (concepts) derived by the authors in group discussions (see section 4.1). For further analysis, relevant characteristics of the approaches proposed in the articles were extracted, with minimal differences depending on the respective sub-category. In common, a short description of the methodology for task allocation, the possibility of parallel task execution by humans and robots, the utilized core criteria for task assignment, and the consideration of algorithms for optimizing the process plan or the task distribution were extracted. These characteristics were developed in discussions between the authors based on the knowledge gained during the literature filtering process.

3.5. Research agenda

Finally, after analyzing and describing the approaches, a reference model for task allocation methods was derived based on the common methodologies of the identified approaches, and the limitations of current approaches were discussed. In addition, research needs were derived and discussed. This responds to the suggestion of (Webster and Watson 2002; Vom Brocke et al. 2009) that literature reviews should lead to a sharper research agenda.

4. Results

From the initial keyword-based database search, a total of 521 articles were found after removing the duplicates. After the filtering process, 24 articles remained with high relevance, for which the performed forward and backward search led to 13 additional articles, thus resulting in a total of 37 records included in this literature review. The entire literature search and filtering procedure is shown in Figure 3. The included articles are presented for each sub-category in Tables 4–8.

Figure 3. Flow chart of literature search and filtering procedure conducted in this review.

Table 3. List of top ten most cited papers (data retrieved from Google Scholar on 10^th February 2022) on task allocation for human-robot collaborative assembly.

Reference	Title	Citations
(Tsarouchi, Makris, and Chryssolouris 2016)	On a human-robot collaboration in an assembly cell	179
(Johannsmeier and Haddadin 2017)	A Hierarchical Human-Robot Interaction-Planning Framework for Task Allocation in Collaborative Industrial Assembly Processes	146
(Schrempf et al. 2005)	A novel approach to proactive human-robot cooperation	95
(Takata and Hirano 2011)	Human and robot allocation method for hybrid assembly systems	94
(Tsarouchi et al. 2017)	On a human–robot workplace design and task allocation system	82
(Ranz, Hummel, and Sihn 2017)	Capability-based task allocation in human-robot collaboration	80
(Hawkins et al. 2013)	Probabilistic human action prediction and wait-sensitive planning for responsive human-robot collaboration	77
(Michalos et al. 2018)	A method for planning human robot shared tasks	66
(Beumelburg 2005)	Fähigkeitsorientierte Montageablaufplanung in der direkten Mensch-Roboter-Kooperation	55
(Müller, Vette, and Mailahn 2016)	Process-oriented Task Assignment for Assembly Processes with Human-robot Interaction	54

Table 4. Characteristics of static suitability assessment-based task allocation approaches.

Reference	Classification of suitability assessment approach	Task planning methodology	Considered parallelism of assembly processes	Optimization of allocated assembly process
(Antonelli and Bruno 2018)	Quantitative capability calculation	Suitability: automatic task classification using decision tree classifier based on task indicators Allocation: logic based on task classification and availability (robot preferred)	sequential & parallel	no
(Beumelburg 2005)	Quantitative capability calculation	Suitability: capability index calculation using criteria catalogue and company objectives Allocation: logic based on capability index and process times Optimization: genetic algorithm to optimize skill level and cycle time	sequential & parallel	yes (optimization of skill level and cycle time)
(Evangelou et al. 2020)	Qualitative suitability assessment	Suitability: manual input of task-resource suitability Allocation: automatic generation of assignment sub-plans (branches) Optimization: branch selection by best utility score using iterative deepening search algorithm	sequential & parallel	yes (optimization of multi-criteria utility value)
(Gjeldum et al. 2021)	Qualitative suitability assessment	Suitability: subjective decision based on matching of capabilities and task requirements Allocation: manual definition of alternative task allocation options Optimization: selection of option using multi-objective ant colony optimization algorithm with criteria weighting by user input	sequential & parallel	yes (optimization of cycle time, investment cost, worker effort, and workspace layout)
(Lee et al. 2022)	Qualitative suitability assessment	Suitability: manual task classification based on task time and safety Allocation: automatic generation of allocation alternatives using search algorithm Optimization: Search algorithm to minimize cycle time	sequential & parallel	yes (optimization of cycle time)
(Liau and Ryu 2020, 2022)	Quantitative capability calculation	Suitability: functional action-resource assignment priority ranking based on skill and part requirements using Analytic Network Process Allocation: logic based on priority ranking and availability Optimization: Genetic algorithm to minimize cycle time	sequential & parallel	yes (optimization of cycle time)
(Malik and Bilberg 2019)	Automation potential assessment	Suitability: automation potential based on criteria catalogue Allocation: logic based on automation potential and availability	sequential & parallel	no
(Müller, Vette, and Mailahn 2016)	Qualitative suitability assessment	Suitability: subjective decision based on pairwise comparison of skills and task requirements Allocation: manual task assignment	sequential & parallel	no
(Amir, Fathi, and Ng 2022)	Qualitative suitability assessment	Suitability: manual suitability assessment based on task time as input Allocation: automatic generation of allocation alternatives during neighborhood search Optimization: Local search-based heuristic algorithm to minimize cycle time and number of operators	sequential & parallel	yes (optimization of cycle time and number of operators)
(Ranz, Hummel, and Sihn 2017)	Quantitative capability calculation	Suitability: knock-out list for unambiguous task assignment and calculation of capability index based on criteria catalogue for remaining tasks Allocation: process plan results from best capability index	sequential & parallel	no
(Schmidbauer et al. 2020)	Quantitative capability calculation	Suitability: task classification using criteria catalogue Allocation: manual task assignment	sequential	no
(Takata and Hirano 2011)	Qualitative suitability assessment	Suitability: qualitative comparison of skills and task requirements Allocation: manual selection of initial allocation pattern Optimization: Genetic algorithm to minimize production costs for future production changes	sequential	yes (optimization of total production costs)
(Yuan et al. 2021)	Automation potential assessment	Suitability: automation potential based on criteria catalogue Allocation: manual input of task allocation alternatives based on automation potential and selection of plan using plan evaluation logic	sequential	no

Table 5. Characteristics of static simulation-supported, capability-based allocation approaches.

Reference	Purpose of simulation	Task planning methodology	Considered parallelism of assembly processes	Assembly process optimization
(Bänziger, Kunz, and Wegener 2018)	Evaluation of different task allocations considering the actual resource movements	Suitability: manual, binary decision Allocation: arbitrary initial allocation Optimization: two-layer optimization of task distribution and task sequence using genetic algorithms	sequential & parallel	yes (optimization of cycle time, ergonomics and movement distance for collision avoidance)
(Heydaryan, Suaza Bedolla, and Belingardi 2018)	Evaluation of different workplace layouts (optimization of robot’s position)	Suitability: binary suitability assessment based on analytic hierarchy process (AHP) Allocation: logic based on hierarchical task analysis (HTA)	sequential	no
(Komenda, Ranz, and Sihn 2019)	Assessment of actual task execution times for different safety modes considering actual resource movements	Suitability: manual decision Allocation: manual input of task allocation alternatives Optimization: calculation of task allocation with minimal cycle time	sequential & parallel	yes (optimization of cycle time)
(Michalos et al. 2018)	Evaluation of different task allocations and workplace layouts	Suitability: binary suitability assessment based on multi-criteria decision-making framework using CAD data Allocation: automatic generation of assignment alternative samples based on suitability constraints Optimization: sample selection by best utility score using depth-first search algorithm	sequential & parallel	yes (optimization of multi-criteria utility value)
(Tsarouchi et al. 2017)	Evaluation of different task allocations and workplace layouts	Suitability: binary suitability assessment based on multi-criteria decision-making framework Allocation: calculation of utility value for given task allocations	sequential & parallel	no

Table 6. Characteristics of reactive and ad-hoc approaches for dynamic task allocation.

Reference	Reactive plan adaptation approach	Task planning methodology	Considered parallelism of assembly processes	Assembly process optimization	Purpose of simulation
(Antonelli and Bruno 2019)	Adaptation of initial task allocation plan (assignment to idle resource in case of delays)	Suitability: automatic task classification using decision tree classifier based on task indicators Allocation: logic based on suitability, availability and operation time (robot preferred)	sequential & parallel	no	no simulation
(Bruno and Antonelli 2018)	Ad-hoc task allocation logic based on task classification (robot preferred)	Suitability: automatic task classification using decision tree classifier based on task indicators Allocation: logic based on task classification and availability (robot preferred)	sequential & parallel	no	no simulation
(Cesta, Orlandini, and Umbrico 2018)	Adaptation of initial task allocation plan (re-planning)	Suitability: manual decision on task classification (human, robot, or human and robot) Allocation: logic based on task classification and operational requirements	n/a	no	no simulation
(Chen et al. 2014)	Ad-hoc task allocation by online application of optimization algorithm	Suitability: n/a Allocation: automatic generation of scheduling alternatives Optimization: genetic algorithm to minimize cycle time and costs	sequential & parallel	yes (optimization of cycle time and cost)	no simulation
(Darvish et al. 2018)	Adaptation of initial task allocation plan (re-planning of remaining path in and-or tree in case of delays)	Suitability: manual decision on capability Allocation: representation of viable alternatives in and-or graph Optimization: calculation and selection of path with minimum cost	sequential & parallel	yes (optimization of multi-criteria utility value)	Validation of the assembly plan feasibility and calculation of utility function
(Faroni et al. 2020)	Adaptation of initial task allocation plan (re-planning in case of delays or human rejection of tasks)	Suitability: manual, binary decision Allocation: representation of partial plan alternatives using timeline-based planning Optimization: search strategy to selects partial plan that optimize workload balancing	sequential & parallel	yes (optimization of workload balancing)	(Validation of proposed approach in simulated use-case)
(Johannsmeier and Haddadin 2017)	Adaptation of initial task allocation plan (re-planning in case of failures)	Suitability: manual, binary decision on capability Allocation: generation of alternative assembly sequence plans based on capability and resource’s task execution time Optimization: divide-and-conquer search method to select plan that optimizes utility function	sequential & parallel	yes (optimization of cycle time and human workload)	(Validation of proposed approach in simulated use-case)
(Makrini et al. 2019)	Ad-hoc task allocation logic based on capability and ergonomics (subsequent human tasks avoided)	Suitability: binary suitability assessment based on mapping of task requirements with resource and object characteristics Allocation: logic based on capability, availability, ergonomics and workload (human workload reduced)	sequential	no	no simulation
(Messeri et al. 2022)	Ad-hoc task allocation logic based on ergonomics (avoid strenuous human tasks)	Suitability: manual, binary input Allocation: logic based on online ergonomics estimation (human’s muscle fatigue reduced)	sequential	no	no simulation
(Nikolakis et al. 2018)	Adaptation of initial task allocation plan (re-planning of remaining operations in case of plan deviations)	Suitability: manual, binary decision Allocation: manual input of alternatives and selection of alternative with highest score based on weighted multi-criteria decision matrix	sequential & parallel	no	no simulation
(Pupa and Secchi 2021)	Adaptation of initial task allocation plan (assignment to idle resource in case of delays or failures)	Suitability: manual, binary decision and input of task execution time Allocation: automatic generation of nominal schedule based on nominal task execution times	sequential & parallel	no	no simulation
(Tsarouchi, Makris, and Chryssolouris 2016)	Adaptation of initial task allocation plan (change assignment in case of unavailable resource)	Suitability: manual, binary decision Allocation: logic based on suitability, availability and time Optimization: sample selection by best utility score using depth-first search algorithm	sequential	yes (optimization of multi-criteria utility value)	no simulation

Table 7. Characteristics of proactive approaches for dynamic task allocation.

Reference	Human tracking system	Intention prediction model	Predicted variable	Robot action selection and execution methodology
(Cramer, Kellens, and Demeester 2021)	RGB-D camera	Dynamic Bayesian network	Assembly sequence desired by human	Adjustable tradeoff between robot task execution and idling (reward-based robot action selection after prediction of assembly sequence conducted by operator; robot prefers actions that are unfavorable for humans)
(Hawkins et al. 2013)	RGB-D camera	dynamic Bayesian network	Human task Remaining human task duration	Tradeoff between robot task execution and idling (minimize overall waiting time by cost function penalizing too early and too late robot task execution)
(Kwon and Hong Suh 2012)	n/a	dynamic Bayesian network	Next human task	Anticipatory robot action selection based on prediction of next required assembly part (minimize overall waiting time)
(H. Liu et al. 2019)	RGB-D camera	Gaussian mixture model	Remaining human motion trajectory	Robot task execution rather than idling (robot selects an available task different from the human task and executes it in parallel; no explicit timing or coordination)
(Schrempf et al. 2005)	n/a	dynamic Bayesian network	Human task	Robot task execution rather than idling (minimize robot idle time by trial-and-error logic for robot task execution in case of prediction uncertainty)
(Zanchettin et al. 2019)	RGB-D camera	higher-order Markov Chain	Human task Remaining human task duration	Robot idling rather than task execution (minimize human waiting time by executing parallel robot tasks only if enough free time is available)

Table 8. Analysis of utilized criteria for task allocation in included references.

Category		Reference	Criteria for task allocation
			resource characteristics	part characteristics	assembly process requirements	task time	cycle time	ergonomics	safety	resource availability	cost or investment	process quality	workload or occupancy	movement length or space utilization
Static	Suitabilityassessment	(Antonelli and Bruno 2018)	✓	✓	✓	✓				✓
		(Beumelburg 2005)	✓	✓	✓		✓	✓			✓	✓
		(Evangelou et al. 2020)	✓	✓	✓		✓	✓	✓	✓				✓
		(Gjeldum et al. 2021)	✓		✓		✓	✓			✓			✓
		(Lee et al. 2022)	✓		✓	✓	✓		✓	✓				✓
		(Liau and Ryu 2020, 2022)	✓	✓	✓		✓	✓		✓
		(Malik and Bilberg 2019)	✓	✓	✓		✓		✓	✓
		(Müller, Vette, and Mailahn 2016)	✓	✓	✓
		(Nourmohammadi, Fathi, and Ng 2022)	✓			✓	✓				✓		✓
		(Ranz, Hummel, and Sihn 2017)	✓	✓	✓		✓	✓			✓	✓
		(Schmidbauer et al. 2020)	✓	✓	✓			✓
		(Takata and Hirano 2011)	✓		✓						✓
		(Yuan et al. 2021)	✓	✓	✓			✓	✓				✓
	Simulation-supported	(Bänziger, Kunz, and Wegener 2018)	✓		✓		✓	✓						✓
		(Heydaryan, Bedolla, and Belingardi 2018)	✓		✓			✓	✓			✓
		(Komenda, Ranz, and Sihn 2019)	✓		✓		✓		✓
		(Michalos et al. 2018)	✓	✓	✓		✓	✓			✓		✓	✓
		(Tsarouchi et al. 2017)	✓	✓	✓		✓	✓			✓			✓
Dynamic	Reactive & ad-hoc	(Antonelli and Bruno 2019)	✓	✓	✓	✓		✓		✓
		(Bruno and Antonelli 2018)	✓	✓	✓	✓				✓
		(Cesta, Orlandini, and Umbrico 2018)	✓		✓				✓
		(Chen et al. 2014)	✓			✓	✓				✓		✓
		(Darvish et al. 2018)	✓		✓	✓
		(Faroni et al. 2020)	✓										✓	✓
		(Johannsmeier and Haddadin 2017)	✓				✓						✓
		(Makrini et al. 2019)	✓	✓	✓			✓
		(Messeri et al. 2022)	✓		✓			✓					✓
		(Nikolakis et al. 2018)	✓	✓	✓		✓			✓
		(Pupa and Secchi 2021)	✓		✓	✓	✓		✓				✓
		(Tsarouchi et al. 2016)	✓	✓	✓	✓	✓			✓
Total number of applying approaches			30	16	26	9	17	14	8	9	8	3	8	7

With the relevant articles included in the review, the distribution of the scientific contributions over the years was analyzed (Figure 4). After a few pioneering publications on the subject, e.g. the dissertation by (Beumelburg 2005) in 2005, the topic has been gaining increasing interest since 2016. Noteworthy are the four relevant publications found for the year 2022, as the database search was conducted in January 2022, indicating that the research field remains highly relevant.

Figure 4. Number of publications per year on task allocation for human-robot collaborative assembly.

As shown in Figure 5, the number of conference papers (19) and journal papers (18) is almost leveled, underlining the identified importance of the topic as well as the advanced development stage of some approaches. For both conferences and journals, only very few articles were published in one and the same medium. Instead, the identified articles were published widely spread in a variety of well-known conferences and journals, mainly dealing with smart manufacturing or robotics. While this underlines the relevance for a broad audience, it also reveals a lack of specialized conferences or special issues dealing with task allocation between humans and robots or, more generally, with collaborative assembly.

Figure 5. Number of conference papers, journal articles, and dissertations.

The included 37 papers were written by a total of 130 contributors, resulting in an average of 3.42 authors per publication. Regarding the countries of publication, researchers from institutes in Italy, Germany, and Greece published the most contributions with a share of 57% (21/37) of all included publications, as shown in Figure 6. This refers not only to the quantity but also to the relevance of the corresponding publications from these countries, which account for 80% of the ten most-cited publications (five from Germany, two from Greece, one each from Italy, Denmark, and the United States). Here, it is noteworthy that only the affiliation of the first author was taken into account when analyzing the number of publications per country. The list of the top ten most cited contributions, with their respective citation count retrieved from Google Scholar on 10^th February 2022, is presented in Table 3.

Figure 6. Number of included publications per country.

4.1. Current state of the art of task allocation in assembly: methods and classification (RQ1)

The problem of task allocation between humans and robots consists of several sub-problems that need to be addressed. First, the decision for each individual process step, whether human, robot, or both together should perform the task, must be made. This decision can be made either based on the isolated process step or by including knowledge about the entire pending process. Second, this leads to questions concerning the chronological order in which the process is executed. The given input information for addressing this issue is typically an assembly process sequence (e.g. assembly priority graph), containing the logical sequence of the overall process. For purely sequential assembly processes, the execution order is given and does not need to be taken into account for task allocation. However, for assembly processes in which the parallel execution of sub-processes is possible, e.g. for several sub-assemblies that are prepared in parallel and finally integrated into the overall component, the execution order is not completely predetermined. Instead, a pool of possible subsequent tasks is available from which to choose. Third, in collaborative assembly, obstructions and collisions between the human and the robot as well as delays or failures during the process runtime can occur. Both affect the feasibility or execution time of the process and should be considered in holistic task allocation approaches. These sub-problems must be taken into account in order to best achieve enterprise objectives (e.g. time, costs, quality, flexibility, ergonomics) and to exploit the potential and advantages of HRC in assembly.

Several approaches have been published to address these issues, among which the authors of this review identified two superordinate categories and two sub-categories for each. As mentioned above, this review paper generally distinguishes between static and dynamic task allocation approaches. Static (or offline) methods tackle the problem of task allocation prior to the actual execution of the assembly, i.e. once the task allocation has been defined for the process, it cannot be changed during execution. On the contrary, the final allocation of tasks in dynamic (or online) approaches takes place during the assembly, allowing the current situation to be taken into account.

Thus, the considered publications and corresponding approaches were classified into this proposed taxonomy as shown in Figure 7. 19 papers suggest static allocation methods, while 18 articles consider dynamic task allocation. Also, the number of approaches that apply optimization algorithms to refine the task schedule is indicated in Figure 7. In the following, the general idea for the allocation logic for each of the individual approaches is presented, starting with the static and followed by the dynamic approaches. Also, the respectively identified sub-categories are explained in the following.

Figure 7. Categorization of task allocation approaches including number of considered publications per category.

4.1.1. Static task allocation

Static approaches typically apply different criteria to calculate the suitability of a resource (human or robot) for certain processes or task types. Here, some approaches are solely based on those prior defined criteria to find a suitable task allocation plan, while other approaches additionally utilize, at a later stage, a simulation to check, refine, or optimize the task allocation plan. Therefore, a classification of static task allocation approaches into suitability assessment-based allocation approaches and simulation-supported, capability-based allocation approaches is proposed.

4.1.1.1. Task allocation based on suitability assessment

As a common baseline of the identified approaches, the suitability of humans and robots is compared with the requirements of the tasks within assembly processes (suitability-check). Then, additional but optional, more general criteria are applied to assess the reasonableness of the allocation to a respective resource (reasonableness-check). Finally, a decision logic is applied that takes the complete assembly process into account and decides on the task allocation of the overall process (task allocation logic). In addition, a few approaches apply optimization methods in a last step to refine the initial task allocation plan (plan optimization). The output is, in each case, a static process plan to be executed in a fixed manner by the resources.

For calculating suitability, a straightforward approach is to rely on human workers for tasks that are too expensive to automate (Antonelli, Astanin, and Bruno 2017). Some approaches follow this idea by assessing the automation potential of individual process steps, e.g. (Malik and Bilberg 2019; Yuan et al. 2021). However, this fundamental idea of relying on humans for hard-to-automate tasks leads to the risk that the human worker is degraded to a stopgap or appendage of the automation (Lotter and Wiendahl 2012), which must be prevented for a successful implementation of HRC, especially with regard to workers’ acceptance.

Therefore, methods have been proposed that conduct a qualitative pair-by-pair comparison of the capabilities of both resources, e.g. the assessment of advantages and disadvantages of both resources for each process step (Müller, Vette, and Mailahn 2016). These methods mainly aim at a fast assessment and creation of a possible process plan, which, however, is most probably not optimal on the one hand and subjective on the other.

To overcome subjectivity, other approaches have been proposed that aim to objectify the decision and mapping of tasks to suitable resources using criteria catalogues in combination with a quantitative rating system, e.g. (Beumelburg 2005; Ranz, Hummel, and Sihn 2017; Schmidbauer et al. 2020) or other quantitative ranking methods for multi-criteria decision analysis, e.g. analytic network process (ANP) (Liau and Ryu 2020) and decision trees (Antonelli and Bruno 2018).

Thus, based on (Ranz, Hummel, and Sihn 2017), methods for the suitability calculation can be distinguished between three fundamental approaches:

1. assessment of automation potential (objective quantification of robot’s suitability),

2. qualitative suitability assessment or comparison of capabilities (subjective, but quick orientation), and

3. calculation of capability indicators (more objective quantification of both resources).

Regarding the task allocation decision logic, the existing approaches either let the human decide on a process plan based on the process-specific suitability information, e.g. (Müller, Vette, and Mailahn 2016; Schmidbauer et al. 2020), or suggest a more defined decision procedure based on the given suitability and additional information, such as availability or time. Here, some approaches preferentially assign tasks to either the human, e.g. (Antonelli and Bruno 2018) or the robot, e.g. (Malik and Bilberg 2019), in case of ambiguous pre-assignments. To further advance towards an optimal static process distribution, some approaches apply optimization methods to refine the process plan in terms of time (Liau and Ryu 2020; Beumelburg 2005; Lee et al. 2022; Amir, Fathi, and Ng 2022; Liau and Ryu 2022), capability (Beumelburg 2005), costs (Takata and Hirano 2011), or multi-criteria utility functions (Evangelou et al. 2020).

A brief description of each analyzed reference within the sub-category ‘task allocation based on suitability assessment’ in terms of their approach for suitability assessment and the task allocation logic applied is provided in Table 4.

4.1.1.2. Quantitative summary

A brief quantitative summary of the characteristics of the approaches presented in Table 4 is provided in the following: Most approaches (11/14, 79%) consider a parallel execution of tasks by humans and robots, while three only allow sequential execution of tasks. About half of the approaches (8/14, 57%) conduct a posterior optimization of the process to minimize, e.g. cycle time or maximize exploited capabilities. Regarding the methodologies utilized within the suitability assessment, most references suggest quantitative methods for the capability calculation of both resources (6/14, 63%). The actual task allocation for the overall process is then conducted either completely manual, using previously collected information as a decision basis (5/14, 36%), or supported by an explicit decision logic (6/14, 43%), or automatic as part of schedule optimization (3/14, 21%).

4.1.1.3. Simulation-supported, capability-based task allocation

The previously presented analytical approaches do not allow, despite the application of optimization algorithms in some publications, a validation of the resulting process prior to process execution. Therefore, an extension by simulation environments have been proposed by a number of references. Approaches within this sub-category are still static task allocation methods, i.e. their output is a fixed assembly process plan. However, the advantage of using a simulation is that the dynamics of the interaction between the worker and the robot in their shared workplace can be considered when evaluating different tasks assignments and schedules (Bänziger, Kunz, and Wegener 2018).

The focus of the identified publications within this sub-category is mostly on the introduction and use of the simulation environment. Thus, less detail is given to the capability assessment or the task distribution logic itself, or more simplistic approaches are used for this, e.g. by using only binary values to represent resource capabilities for task execution, which express whether a resource is capable in principle, but not whether execution by the respective resource is appropriate (Komenda, Ranz, and Sihn 2019; Bänziger, Kunz, and Wegener 2018; Heydaryan, Suaza Bedolla, and Belingardi 2018). The simulation is either utilized to evaluate and compare different assembly task allocation schedules (Michalos et al. 2018; Tsarouchi et al. 2017) or combined with an optimization approach (Komenda, Ranz, and Sihn 2019; Bänziger, Kunz, and Wegener 2018). Also, the workplace layout, e.g. the position of the robot, can be optimized in simulation for a given assembly process allocation (Heydaryan, Suaza Bedolla, and Belingardi 2018; Tsarouchi et al. 2017; Michalos et al. 2018).

A brief description of the different simulation-supported approaches and their characteristics, highlighting in particular the purpose of the simulation, is given in Table 5.

4.1.1.4. Quantitative summary

As follows from Table 5, the simulation is used for workplace layout optimization and visualization (3/5, 60%), and for evaluating additional parameters for task allocation (4/5, 80%), including the estimation of the actual resource movements by using a digital model of both human and robot (2/5, 40%). The digital human model thus enables, on the one hand, the identification of possible collisions or obstructions leading to potential delays during task execution and, on the other hand, the assessment of ergonomics criteria. While some approaches automatically generate alternative task distributions as input for the assessment in the simulation (2/5, 40%), others rely on a predefined input of task allocation alternatives to be evaluated (3/5, 60%).

4.1.2. Dynamic task allocation

As shown in the previous section, static task allocation approaches aim to find a reasonable or even optimal task sequence offline, which is then exactly followed by robots and humans during execution. However, the reliability of the actual execution of an offline task allocation plan is limited due to uncertainties and changes in the workspace as well as unpredictable events and human decisions (Darvish et al. 2018). This might be caused by a wide range of reasons, e.g. due to safety-related evasive movements or speed reductions for collision avoidance, stops caused by unintentional contact with workers, expertise- or health-related performance differences among individual workers, attention-, motivation- and fatigue-related performance variations throughout the day, or restrictions in robot perception (Wang et al. 2019; Chen et al. 2014; Antonelli and Bruno 2019; Faroni et al. 2020). Even though some of these aspects, such as collision avoidance movements, are addressed and estimated by the aforementioned simulation-supported methods, the actual uncertainties introduced by the human in the process (Wang et al. 2019) cannot be modeled. Especially in non-standardized productions that are characterized by small lot sizes, outages are frequent, and the time spent on the individual task can have high variability resulting in offline optimized task assignments to become ineffective and sub-optimal (Antonelli and Bruno 2019).

Therefore, dynamic task allocation approaches have been proposed in the literature to circumvent the problem of under-specification by adapting to a world of change (Hancock and Scallen 1996), aiming to increase flexibility and overall success rate of collaborative assembly processes (Darvish et al. 2018).

The identified references for dynamic task allocation can be classified into two main categories, namely reactive and ad-hoc task allocation and proactive task allocation. Reactive allocation approaches follow a centrally orchestrated and monitored task allocation plan, which is reactively adjusted as changes occur or apply an ad-hoc allocation logic. In contrast, proactive approaches let the human take the lead in the assembly process, and the system then dynamically decides which work steps the assisting robot can perform in anticipation to proactively support the human worker. The following sections describe the two categories in more detail.

4.1.2.1. Reactive and ad-hoc task allocation approaches

Reactive task allocation approaches have a central planning instance in common, which assigns tasks to the human or the robot. Two types of methods have been identified within this sub-category: First, some approaches are based on an initial plan that is dynamically adapted in case of disturbances, e.g. by changing individual assignments (Antonelli and Bruno 2019; Pupa and Secchi 2021; Tsarouchi, Makris, and Chryssolouris 2016), or by triggering a re-planning of all remaining tasks (Faroni et al. 2020; Cesta, Orlandini, and Umbrico 2018; Darvish et al. 2018; Nikolakis et al. 2018; Johannsmeier and Haddadin 2017). In the second type, the allocation is solely based on an ad-hoc decision logic for allocation, which is applied online for assigning each task, e.g. based on the task classification (Makrini et al. 2019; Bruno and Antonelli 2018), or based on human fatigue (Messeri et al. 2022). However, it is worth mentioning that these approaches can merge and overlap, particularly when an ad-hoc logic is applied for task reassignment in case of deviations from the initial plan.

Further, reactive approaches of the identified references differ in terms of decision-making authority. While some approaches are suggestion-based, allowing the human to decide whether or not to perform the assigned task, e.g. (Faroni et al. 2020; Messeri et al. 2022), most follow a command-based methodology, where the decision-making authority belongs to the central task planning system, e.g. (Antonelli and Bruno 2019; Nikolakis et al. 2018).

To adapt and react to changes from plan execution or even to trigger ad-hoc allocation logic, monitoring the process execution to detect the completion of an assembly process step is required. This state information is gathered from the robot controller for the robot’s task execution. For human tasks, the confirmation of task completion can be detected either automatically by monitoring the assembly workplace to track and classify human actions, e.g. (Darvish et al. 2018), or manually by confirming task execution, using gestures or physical, respectively digital buttons, e.g. (Faroni et al. 2020).

In contrast to static task allocation approaches, dynamic approaches are limited in terms of computation time as the final decision on an allocation is not made prior but online during assembly task execution. Therefore, genetic optimization algorithms, such as those applied by (Beumelburg 2005; Liau and Ryu 2020; Takata and Hirano 2011), which require significant computation times until convergence, cannot be applied. Thus, a key requirement of corresponding dynamic algorithms is the need to be computationally efficient. Consequently, the existing reactive and ad-hoc allocation approaches are based on rather simple online decision logics.

A brief description of the reactive and ad-hoc task allocation approaches is provided in Table 6.

4.1.2.2. Quantitative summary

From the included references within this sub-category, the vast majority (8/12, 67%) rely on an initial task allocation plan which is calculated offline and only adapt this plan in the event of plan deviations, e.g. due to delays. The remaining approaches directly apply an ad-hoc logic for task allocation online, following rather simple decision logics (3/12, 25%), or even argue to be able to conduct optimization of the task allocation directly online (1/12, 8%, (Chen et al. 2014)). 42% (5/12) of the considered approaches perform an (offline) optimization of the task allocation to account for additional criteria, such as cycle time or workload, mostly based on tree search algorithms. Different from the approaches discussed in the previous sub-category, here, only one approach (10%) uses a simulation environment within the task planning methodology to calculate the value of a utility function as input for an optimization algorithm. The suitability assessment in this sub-category is rather simple, despite the fact that most approaches rely on an initial plan. However, as the approaches focus on the dynamic online adaptation methodology, the suitability assessment is given less attention. As a result, 75% of the approaches (9/12) rely on a fully manual, qualitative capability assessment, respectively, task classification, or do not even mention the suitability assessment at all. The remaining quantitative suitability calculation methods are mostly based on approaches that were presented in previously published work (17%, 2/12).

4.1.2.3. Proactive task allocation approaches

In each of the three sub-categories presented above, approaches are described in which the planning and distribution of assembly tasks is carried out by a central planning instance. However, the planning of assembly tasks in dynamic environments can also be driven by the human (c.f. (Fiore, Clodic, and Alami 2016)). In such approaches, the human takes on the role of the team leader (H. Liu et al. 2019) and independently performs self-selected assembly tasks, while the robot aims to support the human by proactively selecting tasks that can be performed in parallel as preparatory work for subsequent steps to faster achieve the joint goal of completing the entire assembly process. The idea for such assembly process execution is inspired by human teamwork, as humans are able to act proactively and assist others by predicting their intentions before help is actually requested (Kwon and Hong Suh 2012).

To be able to anticipate the human’s intentions, these proactive task allocation approaches require continuous monitoring of the assembly execution in combination with human movement prediction algorithms, movement interpretation, and task classification (c.f. (Clodic et al. 2017)). The monitored information on the currently executed tasks and the state of the human operator is then utilized for the proactive selection of useful auxiliary actions or process steps based on the task model representation (H. Liu et al. 2019), e.g. based on an assembly priority chart.

Since the intention prediction and action selection need to be performed online and as quickly as possible, a key requirement for the algorithms is real-time capability. Another challenge is the uncertainty handling and temporal synchronization of the task execution. On the one hand, the waiting time should be minimized by quickly deciding on the execution of an auxiliary action, while, on the other hand, a high probability of correctly predicting the intended human action and its temporal end is desirable. To address these issues of timing and uncertainty handling, the approaches for proactive task allocation suggest different degrees of proactivity for the robot’s behavior. Two extremes were identified: Either the robot’s maxim is to be minimally disruptive to humans, forcing the robot to idle in the case of a high probability of causing human waiting times, e.g. (Zanchettin et al. 2019), or the robot is more active and prefers to perform a wrong task rather than idling, despite possible human disruption. For the latter, i.e. performing robot tasks even under high uncertainty, (Schrempf et al. 2005) propose a trial-and-error logic for supportive tasks and (H. Liu et al. 2019) suggest to select tasks that do not affect the human. Also, tradeoffs between robot idling and execution have been proposed, e.g. based on a cost function that penalizes both extremes (Hawkins et al. 2013) or by an adjustable reward function (Cramer, Kellens, and Demeester 2021).

For estimating human intentions, the identified references utilize inference mechanisms in combination with probabilistic statistical models, such as dynamic temporal Bayesian networks (DBN), e.g. (Kwon and Hong Suh 2012; Schrempf et al. 2005; Hawkins et al. 2013), higher-order Markov Chains, e.g. (Zanchettin et al. 2019), or Gaussian mixture models (GMM), e.g. (H. Liu et al. 2019). For human motion tracking, the referenced approaches mostly utilize a depth-image camera (RGB-D camera), e.g. (Hawkins et al. 2013; H. Liu et al. 2019; Zanchettin et al. 2019; Cramer, Kellens, and Demeester 2021). This is consistent with the findings from (Arents et al. 2021), who identified the use of 3D cameras in almost 50% of their analyzed publications on HRC. As human motion prediction in HRC is not the focus of this paper, readers interested in this topic are referred to the reviews from (Rudenko et al. 2020; Mao et al. 2013) and the classification of movement prediction methods for HRC presented in (Galin and Meshcheryakov 2019).

As outlined above, the identified approaches within this sub-category – in contrast to the previous sub-categories – mainly focus on the methodologies for human movement prediction and the triggering of robot tasks, while the implications to task planning of collaborative assembly processes are only partially addressed. As a result, these approaches neither apply capability calculation criteria, nor suitability assessment methods, nor optimization algorithms. Therefore, the columns in the overview on the approaches in Table 7 are widely changed, showing the motion tracking system, the intention prediction model, the inferred prediction variable, and the action selection method.

In general, the advantage of proactive task allocation approaches is a higher degree of human decision-making and flexibility in assembly execution, implying that the work can be performed in a less monotonous manner. However, this can come at the expense of lower process efficiency, higher unpredictability of cycle times, and a greater performance divergence among different workstations, depending on the human operator. Furthermore, as the robot could perform either unnecessary or incorrectly predicted actions based on presumed human intentions, the productive time of the robot is reduced. This results in a worse and more difficult to calculate return on investment for the robot. Moreover, if the prediction of the correct intention is determined too late to perform a supporting robot action in time, the human may have to wait, which in turn might also reduce the overall productivity.

4.1.2.4. Quantitative summary

Apart from the different utilized probabilistic prediction models discussed above, the included references also differ in terms of the feature which is predicted by the prediction model, as summarized in Table 7: The current human task is predicted by 50% (3/6) of included approaches, partly together with the remaining task duration (33%, 2/6). Further, there is one approach each that predicts the remaining human motion trajectory, anticipates the human’s next task based on the task model, or even predicts the complete assembly sequence that the human is most likely to follow. Accordingly, the methodologies for robot action selection and coordination differ with respect to the consideration of time as only such approaches that predict the remaining human task duration or anticipate future tasks can conduct a temporal scheduling, e.g. to minimize the overall waiting time (33%, 2/6) or the human waiting time (17%, 1/6).

4.2. Criteria for allocating assembly tasks (RQ2)

Based on the investigation and classification of the state of the art approaches for task allocation, the criteria that are utilized in the included references as a decision basis for allocating assembly tasks to humans or robots were analyzed. As the sub-category of proactive task allocation approaches does not apply an analytic assessment methodology for analyzing suitable task distribution and the included references do not provide information about applied criteria to initially decide about potential assistive tasks prior to online support anticipation, this sub-category is excluded from the criteria analysis.

From the included references, the utilized criteria were extracted and arranged into twelve criteria groups. Table 8 shows the resulting detailed analysis of utilized criteria in each considered approach. In the following listing, these criteria are briefly described, and some exemplary specific parameters are given. For the quantitative calculation of the individual criteria, equations are presented in various publications; for those, the reader is referred to the references (Tsarouchi et al. 2017; Michalos et al. 2018; Malik and Bilberg 2019; Robotiq 2018; Evangelou et al. 2020).

4.2.1. Resource characteristics

The general properties and limitations of both resources (human and robot) are described by this criterion to derive and assess the basic eligibility for the execution of operations. For the robot, these properties usually result from the data sheet of the base system in conjunction with the specifications of the end-effector. Exemplary parameters are the maximum robot payload, respectively the maximum acceptable human load, reachability, speed, force exertion, equipped gripper type, gripping force, or precision (c.f. (Nikolakis et al. 2018; Michalos et al. 2018)). For human characteristics, an interesting idea has been proposed by (Johannsmeier and Haddadin 2017), who suggest the use of personal-specific, individualized parameters for the human worker in the form of a worker profile, using properties such as attention level, general experience level, or reliability that are mapped to a cost function or assigned as variables to the respective characteristics.

4.2.2. Part characteristics

The criterion of part characteristics comprises specific properties of the part to be assembled and the individual components required for it, which in turn place derivable requirements on the handling or joining of the objects. Exemplary parameters are the ability of parts and components to be grasped, which are affected by physical characteristics such as the components’ size and thickness, tolerances, weight, shape, stability, or sensitivity as described in more detail in (Malik and Bilberg 2019).

4.2.3. Assembly task and process requirements

The key requirements for executing an assembly task in HRC are derived from the assembly task itself. Thus, on the one hand, this criterion covers characteristics arising from the process design, the arrangement of the workstation, or the provision of materials. Subsequently, relevant parameters are, e.g. the task type (joining, handling, controlling, fitting, auxiliary operation), the type of part feeding, spatial conditions such as accessibility of the joining location, the mounting direction, the insertion resistance or required force, precision, the need for holding down after joining, or the type of fastening such as screwing, gluing, bending, and riveting (c.f. (Malik and Bilberg 2019)). On the other hand, the task precedence constraints of the overall process – as the fundamental information required for creating a suitable task schedule – are covered by this group of criteria.

4.2.4. Task time

In addition to the general capability, the execution time of an individual task through the eligible resources is an evident criterion for deciding whether humans or robots should execute a task – the resource that executes a task faster is an obvious candidate for the execution of the task. However, for a more profound task allocation, the consideration of the whole process is always decisive because in parallel branches a longer execution time of individual tasks does not necessarily imply a longer total time, but on the contrary, a parallelization is enabled by taking over tasks by the other resource. Nevertheless, the execution time of tasks can provide a good starting point for deciding on the reasonableness of a task assignment, especially regarding the economic efficiency as well as for time-critical processes.

4.2.5. Cycle time

The cycle time criterion describes the total time that is required for one execution of the entire assembly process and thus represents the more important criterion for the overall process compared to the task time. In contrast to the task time, this criterion cannot be examined in isolation, i.e. per process step, during the suitability assessment but must always be considered in the context of the overall assembly process. Accordingly, this criterion is especially taken into account by approaches that perform an optimization of the overall process in terms of an optimization goal or at least an aspect in the optimization. In addition to cycle time, the time spent on non-value-adding activities can also be taken into account, e.g. (Evangelou et al. 2020).

4.2.6. Ergonomics

The criterion of ergonomics takes into account the human during the assembly process and refers both to physiological ergonomics through ergonomic posture or avoidance of permanent strain, and to environmental conditions potentially causing health risks, e.g. through hazardous substances, dust, high temperatures or noise pollution. The approaches to task allocation examined in this review mostly consider aspects related to posture and physical strain. For instance, regarding the human body posture, parameters such as upper body inclination, lateral shoulder inclination, twisting of the upper body, arm and wrist tilting, or overhead work (c.f. (Petzoldt et al. 2020a)) could be analyzed. For this, also ergonomics assessment methods (see, e.g. (Lowe, Dempsey, and Jones 2019; Heidl, Boespflug, and Wakula 2018)) can be applied to quantify the ergonomic influence of individual process steps. The aspect of cognitive ergonomics and mental stress is addressed as part of the workload criterion described below.

4.2.7. Safety

As outlined in section 2, the safety aspect is crucial for HRC. Safety considerations can also be a decision factor when deciding on a task assignment. Here, for instance, the risk of injury or the probability of collisions as well as the required safety mode and sensor technologies are exemplary parameters to consider when deciding whether the execution of process steps is preferable by one of the resources.

4.2.8. Resource availability

The binary question of whether a resource is available at all for executing a task represents a basic prerequisite for assigning a task to a resource. For the initial process planning, this is partly implicitly considered by a task planner. However, the approaches considered here are those that explicitly check the availability of the resource, e.g. from the previous process step as part of the proposed decision logic for task allocation – either offline, e.g. (Malik and Bilberg 2019; Liau and Ryu 2020) or online, e.g. (Bruno and Antonelli 2018). Especially in dynamic task allocation approaches, this criterion is important as the non-availability of a resource mostly indicates a delay or a plan deviation, resulting in the need to initiate a plan adjustment or re-planning. Apart from the availability of the human or robot resources, the availability of a limited quantity of required tools has also been considered as an additional constraint for task allocation, e.g. in (Lee et al. 2022).

4.2.9. Cost or investment

Due to the high cost pressure in assembly, costs are always an important target criterion. The cost criterion can consider both ongoing costs during production and required investment costs. The latter is primarily assessed by static task allocation approaches either in terms of additionally required equipment such as camera systems, safety functions, or end effectors for the execution of tasks, e.g. (Beumelburg 2005), or in combination with the optimization of the arrangement, design and quantity structure at the workplace itself (Michalos et al. 2018). Ongoing costs are composed, e.g. of labor costs (salary per hour, with total worker hours required for the operations estimated, e.g. by MTM), costs for consumables, energy costs (energy consumption and peak power), and maintenance costs. Together, the total production costs calculated as the sum of the labor and purchase costs (Takata and Hirano 2011) can also be considered. When investigating assembly lines instead of single workstations, some approaches, e.g. (Nourmohammadi, Fathi, and Ng 2022; Gjeldum et al. 2021), also aim to reduce the required number of human operators for cost savings.

4.2.10. Process quality

High quality of the assembly process is a critical requirement. This can be assessed by the work quality or by the process reliability. First, for the quality of work, the error rate, reject or complaint rate can be used. In particular, the error rate can be recorded per process step and can directly be related to the assignment of tasks. For the task allocation to humans or robots, however, this is difficult to measure online during the assembly process, which is why it is usually limited to static task allocation approaches based on statistical analysis in advance. Second, the process reliability can be determined qualitatively on the basis of the process design and the resource characteristics, e.g. the robot allows a consistent quality even for monotonous tasks. Quantification can be achieved, for instance, by better-equal-worse scoring (Beumelburg 2005).

4.2.11. Workload or occupancy

This criterion summarizes several parameters that evaluate the workload and division of labor during the hybrid assembly process. The goal is not only to avoid unproductive time, especially for humans due to comparatively high labor costs, but also to positively affect different aspects of cognitive ergonomics, such as the avoidance of stress or psychological burden caused by the collaborative process execution. Exemplary parameters that are applied in the included references are the human waiting time (i.e. idle time), e.g. (Chen et al. 2014; Pupa and Secchi 2021), the human workload, e.g. (Johannsmeier and Haddadin 2017), the occupancy rate of the robot, e.g. (Michalos et al. 2018), or the balancing of the workload, e.g. (Faroni et al. 2020).

4.2.12. Movement length or space utilization

Finally, this criterion considers the necessary movement of the robot for executing a task. The background to the idea is that such tasks, in which the robot must either move over large distances or perform complex avoidance trajectories due to the simultaneous presence of the human, should rather be avoided. The reason for this is the resulting impact on time as well as energy costs. This criterion, i.e. the robot path length or movement distance, is only considered by simulation-based approaches. For the sake of completeness, the spatial requirements, i.e. the space utilization, should also be mentioned as a parameter. However, this parameter is exclusively considered by simulation-based approaches, which include the optimization of the workplace layout in addition to the task allocation.

Figure 8 shows the percentage of usage of the criteria groups in the studied approaches. This listing may provide both researchers and practitioners an initial indication of relevant criteria that are important in the allocation of assembly tasks in collaborative assembly scenarios. Resource characteristics (100%; 30/30) and process requirements (87%; 26/30) constitute the basic criteria being applied by almost all approaches, partly in conjunction with the assembly part characteristics (53%; 16/30). Furthermore, the time criterion is considered by nearly all approaches (70%, 21/30), either in terms of cycle time, mostly used by static task allocation approaches, or in terms of task time (30%, 10/30), especially applied by dynamic task allocation methods. Moreover, ergonomic aspects are also a relatively common criterion (47%, 14/30), which seems appropriate given the potential of HRC to relieve the physical strain on employees. The other criteria groups are less frequently considered.

Figure 8. Proportion of criteria applied for deciding about task allocation.

Here, it is important to emphasize that all approaches do not rely on a single criteria group but propose either a multi-criteria decision logic or a utility function for optimization, which combines various criteria. In both cases, a weighting of the individual criteria is significant to reach the company’s goals and exploit the potential of HRC. However, in the investigated approaches, both the weighting and the selection of the criteria themselves are rarely validated in practice or methodically determined through the analysis of standards or literature. Nevertheless, a few approaches, e.g. (Beumelburg 2005), do conduct expert interviews to prioritize the criteria.

In addition, it was observed that although dynamic task allocation approaches apply similar criteria to static approaches, they are often considered more superficially. Moreover, it is worth mentioning that most of these criteria are considered offline for initial plan generation and initial suitability decision, while only the aspects of availability, time, and to some extent, ergonomics are monitored online to detect and respond to delays or unexpected events.

5. Discussion

5.1. Reference model

Based on the reviewed literature and the detailed analysis of the individual task allocation approaches as well as the flowcharts presented in (Tsarouchi, Makris, and Chryssolouris 2016; Antonelli and Bruno 2019; Johannsmeier and Haddadin 2017; Michalos et al. 2018; Gjeldum et al. 2021), a reference model for task allocation in human-robot collaborative assembly was developed, which combines the proposed methodologies and procedures explained in section 4.1 into a common framework model. Within the model, the three sub-categories – static task allocation based on suitability assessment, simulation-supported static task allocation, and reactive and ad-hoc dynamic task allocation – are covered. Proactive task allocation approaches were not integrated due to their fundamentally different approach, as explained above. The developed reference model for task allocation in human-robot collaborative assembly is shown in Figure 9. Depending on the intended type of task allocation, it consists of up to seven general process steps, namely:

Figure 9. Reference model for task allocation in human-robot collaborative assembly.

1. Input criteria provision, where the set of criteria (see section 4.2 and Table 8) is collected or recorded, and an interface for accessing the required criteria for the task allocation approach is provided.

2. Suitability assessment, which assesses the basic capability of the resources for conducting each respective task based on process, product, and resource data, which is then encoded, e.g. as a capability score, a binary suitability, or an automation potential (c.f., section 4.1.1, ‘task allocation based on suitability assessment’; and column ‘suitability assessment approach’ in Table 4).

3. Task allocation methodology, which first conducts the static task assignment based on both the capability assessment to identify non-variable and variable tasks, and additional criteria to rate the meaningfulness of an assignment in case of variable task assignments. Next, a static assembly process schedule is created by utilizing the input information (c.f., section 4.1.1; and keyword ‘allocation’ in the ‘task planning methodology’ column of Table 4-Table 6).

4. Process optimization, which aims to optimize the initial schedule by considering the additional criteria derived from the planning objectives as optimization goals of the optimization algorithm (c.f., section 4.1.1; and column ‘optimization of allocation assembly process’ in Table 4-Table 6).

5. Process evaluation in simulation, which can be used to validate the feasibility of a task allocation schedule, to compare and select from alternatives schedules, or to calculate process schedule evaluation parameters as an input to the process optimization algorithm from step 4 (c.f., section 4.1.1, ‘Simulation-supported, capability-based task allocation’; and column ‘purpose of simulation’ in Table 5).

6. Process implementation and execution, where the created initial task allocation process schedule is finally implemented at the assembly system as well as executed by humans and robots using an execution control system. For static task allocation, the implementation of a fixed robot program is sufficient, while for dynamic task allocation, more flexible or hierarchical control architectures are required (for the latter, see section 4.1.1, ‘Reactive and ad-hoc task allocation approaches’; and section 5.3).

7. Execution monitoring and feedback loop, which observes the progress of the process by means of human feedback or tracking and robot state monitoring, and checks for plan deviations or failures to initiate a plan adaptation for dynamic reassignment of tasks (c.f., section 4.1.1, ‘Reactive and ad-hoc task allocation approaches’; and column ‘reactive plan adaptation approach’ in Table 6).

The model is intended to help developers from both research and practice as a kind of blueprint to design a first system architecture for collaborative human-robot task allocation in assembly. As shown in Figure 9, static task allocation approaches result in the creation of a static schedule (either manually created, optimized, or tested in simulation), i.e. these require performing steps 1–5 at most. This fixed schedule can then be deployed by a system integrator (step 6). Reactive, dynamic task allocation approaches are based on the generated initial process schedule, which is initially transferred to the process execution controller either manually or automatically, but extend the execution control with execution monitoring and rescheduling functions (steps 1–7). In contrast, dynamic ad-hoc approaches directly implement the overall task allocation logic in the real-time process execution control and therefore do not require an initial schedule, i.e. steps 1, 2, and 6–7, with an extension of step 6 by an online task allocation methodology.

5.2. Process implications and assembly scenarios

Within this review, a categorization of task allocation methods into static and dynamic approaches was proposed. Given the developed reference model, this classification seems reasonable, as the respective approaches address a different horizon of consideration and processing steps regarding the entire planning and implementation process, which is either limited to the period before process execution or includes it.

A detailed, qualitative comparison of the different approaches to task allocation that can be implemented with the proposed reference model is shown in Table 9. It presents the characteristics and effects of the different task allocation solutions in terms of the assembly process quality and flexibility as well as human factors, and provides an initial assessment of suitable assembly scenarios and implementation efforts based on the authors’ assessments.

Table 9. Qualitative comparison of the different approaches to task allocation and correlation with industrial assembly scenarios. A qualitative assessment of the listed characteristics is indicated by the following symbols: + (positive characteristic), ○ (neutral), and − (negative).

		Static task allocation			Dynamic task allocation
		Task allocation based on suitability assessment	Task allocation with optimization of process schedule	Simulation-supported task allocation	Reactive task allocation	Ad-hoc task allocation
Features & quality of taskallocation solution	Process quality	○ Quality of process schedule depends on experience of planner, respectively the effort for testing and optimization during deployment	+ Globally optimized process schedule with minimized cycle times	+ Process schedule validated before deployment + Process schedule without obstructive movements between human & robot ○ Combination with optimization algorithm possible	○ Process quality depends on prior static task planning method	+ Suitable process schedule, though not optimal + High concurrent activity time - Varying cycle times - Risk of human errors caused by inattentive workers due to varying work sequences
	Flexibility	- No reaction to delays or outages of robot (delays lead to interruption of entire process)	- No reaction to delays or outages of robot (delays lead to interruption of entire process)	- No reaction to delays or outages of robot (delays lead to interruption of entire process)	+ Reaction and partialcompensation of delays + Reaction to outages of robot	+ Reaction and partial compensation of delays + Reaction to outages of robot
	Adaptability to individual worker	- No adaptation to individual worker	- No adaptation to individual worker	+ Simulative modeling of individual worker performance enables creation of worker-specific, static process schedules	○ Limited adaptation to individual worker	+ Dynamic adaptation to performance of individual worker + Dynamic adaptation to human performance fluctuations over shift
	Human factors	- Risk of monotonous work - Risk of human stress caused by robot waiting times	- Risk of monotonous work - Risk of human stress caused by robot waiting times	- Risk of monotonous work - Risk of human stress caused by robot waiting times	+ Reduction of human stress due to flexible takeover of tasks in case of delays - Risk of monotonous work in case of processes with few deviations	+ Varied work processes reduce monotony + Low level of human stress due to adaptive task allocation + High HRC fluency - Higher cognitive workload
Production scenarios	Recommended assembly scenarios	○ Low to medium-volume production ○ Highly standardized assembly processes with low variations in task execution times and infrequent occurrence of delays	○ High volume production ○ Highly standardized assembly processes with low variations in task execution times and infrequent occurrence of delays	○ High volume production ○ Highly standardized assembly processes with low variations in task execution times and infrequent occurrence of delays	○ Medium to high-volume production ○ Certain flexibility in assembly process execution order ○ Processes with low variation of task times, but regular occurrence of delays	○ Low to medium-volume production ○ High flexibility in assembly process execution order (e.g. high parallelism) ○ High variance in task times
	Recommended workspace setup	○ Separated workspaces for human & robot	○ Separated workspaces for human & robot	○ Shared workspaces for human & robot	○ Shared workspaces for human & robot	○ Shared workspaces for human & robot
	Effort forimplementation	○ High manual effort for detailed process planning ○ Low effort for solution development ○ Low to high effort for integration, depending on testing in separate or shared workspace and the objective regarding iterative schedule improvement	○ Low manual effort for process planning ○ High effort for development of optimization algorithm and cost function ○ Low effort for integration in separated workspaces; medium effort for testing in case of integration in shared workspaces	○ Medium manual effort for draft planning of eligible processes ○ High effort for workspace and resource modeling as well as simulation ○ effort for integration	○ Effort for initial process planning depending on prior static task allocation method ○ High effort for development of process control, progress monitoring system, and human interaction system ○ High effort for integration due to creation of flexible executable tasks for both resources as well as the installation of process tracking and notification system	○ Low effort for process planning ○ High effort for development of process control, progress monitoring system, and human interaction system ○ High effort for integration due to creation of all the flexible executable processes for both resources as well as the installation of process tracking and notification system
		○ High effort for process adaptations	○ Low to medium effort for process adaptations	○ High effort for process adaptations	○ Effort for process adaptations depending on prior static method	○ Low to medium effort for process adaptations

Regarding the potential industrial applications of both categories of task allocation approaches, static methods have the advantage of not requiring additional technology for the workplace except the robot and its controller with the allocation schedule implemented once for each process. In contrast, dynamic task allocation requires additional sensors for human tracking and powerful computing units for real-time processing methods as well as intuitive, active instructions on the tasks to be perfomed, e.g. by means of worker assistance systems with information projection onto the assembly worktop. In addition, dynamic approaches need to cope with different types of uncertainties (see section 5.3) and – in the case of a continuous monitoring of human activities – take into account data privacy aspects into account that might hamper industrialization.

Therefore, static task allocation approaches are preferable for processes with larger lot sizes and well-standardized processes in which only few plan deviations are to be expected (c.f. Table 9). In such application scenarios, the high initial effort for approaches that perform process optimization or process validation in simulation is also justified. For static process schedules, manual verification and formal approval can also be conducted prior to implementation to address potentially required quality assurance processes. This is, for instance, relevant in highly standardized industry sectors such as the automobile industry.

However, with regard to less standardized assembly processes or processes with regular occurrence of delays, it is advantageous to have the possibility to react to outages and disturbances by dynamically changing the task distribution during execution instead of having a fixed, albeit initially optimized, schedule. Thus, dynamic task allocation approaches are preferable in such use cases. The decision between reactive and ad-hoc task allocation should be made primarily depending on the permissible variability between different assembly process executions and the desired adaptability to the individual worker (c.f. Table 9).

5.3. Issues and approaches from the wider research field of human-robot interaction

The reviewed and analysed approaches to task allocation in HRC are closely related to research on combined task and motion planning (TAMP) from the wider field of human-robot interaction. Approaches to TAMP aim to combine task planning, i.e. the question of which task to perform, with motion planning, i.e. the calculation of how the motion to complete the task is to be performed (Mansouri, Pecora, and Schüller 2021). A key challenge to TAMP is that respective software tools operate on different representations (discrete and continuous), which makes it hard to link them (Waldhart, Gharbi, and Alami 2016). Therefore, many approaches to TAMP suggest hierarchical control architectures with higher-level task planning and lower-level motion planning (Waldhart, Gharbi, and Alami 2016). On the motion planning level, for instance, human-aware motion planning (Mainprice et al. 2011; Sisbot and Alami 2012) can be applied in human-robot interaction scenarios, which aims to improve HRC fluency, safety, and satisfaction (Lasota and Shah 2015) by choosing trajectories that minimize disturbances towards the human. For task planning, various methods, such as timeline-based task planning (Faroni et al. 2020) or hierarchical agent-based task planning (also referred to as human-aware task planning) (De Silva, Lallement, and Alami 2015; Waldhart, Gharbi, and Alami 2016; Lemaignan et al. 2017; Alami 2013; Fiore, Clodic, and Alami 2016), were suggested. A generic interface for combining both off-the-shelf task planners and motion planners was proposed by (Srivastava et al. 2014). Further, (Chongjie and Shah 2016) present an algorithm to optimize the high-level task planning based on information about the lower-level action and motion selection. For more details on TAMP, the reader is referred to recent review publications that explain current approaches and discuss applications, existing challenges, and limitations of TAMP (Mansouri, Pecora, and Schüller 2021; Garrett et al. 2021 K. Zhang et al. 2022).

Referring to the reviewed literature on task allocation for HRC in assembly, the analyzed approaches in this article mainly consider the aspect of task scheduling and allocation for HRC. However, a dedicated lower-level motion planning, as included in TAMP, is only considered in a few approaches, c.f. (Faroni et al. 2020; Pupa and Secchi 2021; Cesta, Orlandini, and Umbrico 2018). Instead, the trajectories to be executed per task are mostly predefined. The advantage of a dedicated lower-layer for motion planning is, however, that the collaboration dynamics and uncertainties, resulting from the coexistence of the robot and the human in the workspace (Mukherjee et al. 2022; Faroni et al. 2020), can be explicitly addressed. In general, human-robot collaboration needs to tackle different kinds of uncertainties (c.f., (Askarpour et al. 2019)):

• temporal uncertainty, i.e. in response to variations in execution time.

• behavioral uncertainty, including both, intrinsic and interactive uncertainty of human motion. While the intrinsic uncertainty of human motion does not depend on the robot’s state, the interactive uncertainty describes the uncertainty of human’s behavior in reaction to the robot.

• goal uncertainty, describing human uncertainty to perform the required action or to follow the instructions correctly.

Temporal uncertainty is mostly related to task planning and scheduling and is already implicit in most approaches to dynamic task allocation. Future research could focus on incorporating the consideration of behavioral and goal uncertainty into task allocation solutions. Continuous monitoring of both the workplace and the human actions as well as active modeling of uncertainties could contribute to a better management of those uncertainties. In this context, various approaches might be integrated that either model human bevavior (Askarpour et al. 2019; Pöhler, Umlauft, and Hirche 2019), learn its policy (Nikolaidis et al. 2017; Pandya and Liu 2022), or predict human motions (Z. Liu et al. 2019) (see also the analyzed approaches to proactive task allocation). For camera-based human motion detection, (Sajedi et al. 2022) propose an approach to quantify the uncertainty of detection and suggest incorporating uncertainty more transparently in approaches to HRC, e.g. by providing warning signals in case of estimated high uncertainty. This would also contribute to better joint attention and action understanding, which are among the key elements in human-robot interaction (Clodic et al. 2017).

Nevertheless, for the practical implementation of task allocation solutions, a conscious decision should be made about the aspects of uncertainty that need to be considered for each application (Mansouri, Pecora, and Schüller 2021).

Therefore, on the one hand, the explicit consideration and management of those different kinds of uncertainties for task allocation, and, on the other hand, the combination of all aspects – task planning, scheduling, allocation, and (human-aware) motion planning (Mansouri, Pecora, and Schüller 2021) – into an integrated framework are promising research directions towards the design of advanced task allocation solutions with improved effectiveness and reliability.

In this context, an interesting work towards the integration of TAMP and HRC task allocation is proposed by (Pellegrinelli et al. 2017) for the task of (un-)loading a multi-fixturing system. The approach further considers temporal uncertainty by providing a set of trajectories for different human occupancy volumes of the workstation for each task, which constitute an interval for the respective robot execution time. The higher-level task scheduler considers this time interval during scheduling when a human task shall be executed in parallel. In addition, replanning is triggered in case of longer human task durations.

5.4. Challenges and research needs

In addition to the challenges discussed in the previous section 5.3, i.e. addressing uncertainty and exploiting the potentials arising by combining task planning, scheduling, allocation, and human-aware motion planning into an integrated framework, further research needs for task allocation in HRC were identified, which are outlined in the following:

When optimizing the task distribution based on a cost function using an optimization algorithm, different criteria have been combined into metrics as detailed in section 4.2. However, a profound quantification of capabilities and costs of ergonomic impacts for the worker is difficult, and its combination with well-quantifiable factors, such as cost or time, remains a major research challenge (Bänziger, Kunz, and Wegener 2018; Müller, Vette, and Mailahn 2016). According to (Ranz, Hummel, and Sihn 2017), approaches often tend to generalize the capabilities that humans and robots obtain. Especially for human operators, capabilities are in progressive change and depend on the worker-individual skillset, their age, as well as their actual constitution (Ranz, Hummel, and Sihn 2017). Dynamic task allocation approaches have the potential to account for this aspect by implicitly reacting to changes in human performance with adapted task distributions.

Similarly, as already stated in section 4.2, the weighting of individual criteria among each other should be further investigated. So far, literature is not concurring regarding the most suitable criteria to decide on a task allocation. Instead, the included references partly argue differently either for or against, for example, maximizing robot workload, conducting workload balancing, or minimizing human idle time. Here, systematic evaluation studies to investigate the effects and implications of different options and influencing variables constitute an open research question (see, e.g. (Faccio, Bottin, and Rosati 2019) for an investigation of the effects of varying degrees of collaboration, shares of tasks for human and robot, and product complexities for different assembly scenarios). Based on this, assembly scenario-dependent recommendations or methods to automatically adjust applied metrics in response to market demands could be derived.

Besides the applied criteria for task allocation itself, the authors see a need for well-standardized metrics and benchmarks for performance evaluation for HRC in general and specifically for comparing different task allocation approaches. As a starting point, the reader is referred to the overviews from (Marvel et al. 2020; Murphy and Debra Schreckenghost 2013).

Furthermore, the vast majority of the presented approaches in the analyzed references are evaluated in laboratory scenarios and are not yet mature technologies in operational industrial use. This finding is similar to the challenge identified by (Hentout et al. 2019), who state that a large gap exists between the conducted research and the deployment effort in real industrial environments to bring laboratory-based collaborative robot technologies into smart factories. Accordingly, the implementation of both static and dynamic task allocation approaches in practice and the investigation of the resulting implications remains a research need. Also, long-time studies should be conducted to investigate aspects of acceptance, productivity, and responsiveness to frequent variant or model changes.

Moreover, safety is a crucial aspect and an ongoing research challenge related to HRC (Arents et al. 2021). However, as shown in section 4.2, only a few approaches consider safety as an aspect for allocating tasks and seldomly address safety implications for implementation. Also, when implementing task allocation methods in practical applications, it is necessary to ensure that collaboration remains safe, especially during reactive task changes as well as under uncertainty (c.f. section 5.3). Hence, the certification of such applications also remains a future challenge.

Besides the technological part, worker acceptance is a key concern for enabling a successful implementation. A participatory design by involving workers in the implementation process as well as the inclusion of workers into parts of the decision-making process of task allocation could be essential factors for promoting acceptance (Schmidbauer et al. 2020; Tausch and Kluge 2020). The organization of the team either as equal partners or as a leader-assistant setup, as well as the way of assigning tasks based on either suggestions (human with decision-making authority) or commands (only system with decision-making authority), therefore needs to be explored in future work (c.f. (Fiore, Clodic, and Alami 2016)).

Finally, the authors see great potential in expanding current task allocation approaches with further assembly-related technologies. To improve practical usability, the connection of task allocation approaches and informational worker assistance systems that guide operative workers through the assembly process is an important component. Currently, such assembly assistance systems mainly focus on fully manual assembly processes (c.f. (Petzoldt et al. 2020b)). The representation of collaborative processes in such systems in conjunction with intelligent and dynamic task allocation methodologies would provide the operator with the necessary information about both the collaborative process and the precise work instructions in a familiar graphical environment. Thus, both physical and cognitive worker support could be addressed in a combined framework. Apart from the integration itself, the related research question is how and by what modality the human worker should be informed about the next task assigned or possible changes in the plan.

In addition, the integration of intuitive teaching methods to further automatize the entire process could be beneficial. For example, programming by demonstration methods for robot process implementation or teaching by demonstration methods to automatically capture the task sequence or even an initial task allocation schedule constitute a future research direction.

Finally, the motion prediction algorithms applied in the proactive task allocation approaches reviewed in this paper can also be utilized and exploited in more structured architectures, similar to that of the proposed reference model. For instance, rather than just reacting to disturbances, the system could anticipate potential disruptions to avoid delays by preemptively initiating supportive actions that prevent such delays before they occur.

6. Conclusion

This paper conducted a systematic literature review on task allocation for human-robot collaboration in assembly. Human-robot collaboration is attracting increasing interest in both research and practice, with the question of how to distribute processes most effectively posing a significant challenge to a successful breakthrough of collaborative assembly in industry. Nevertheless, while many related reviews examine the general topic of human-robot collaboration, focusing on aspects such as safety, robot programming, communication and interaction, human factors, or applications, only one review was found that explicitly focused on task planning and allocation. However, this review was published six years ago and was not conducted as a systematic literature review. Therefore, this paper intended to address this research gap by focusing on task allocation in assembly as the major application area of human-robot collaboration.

As part of this research, 521 publications resulting from the initial database search as well as the forward-backward search were analyzed, and 37 relevant references were identified that present methodologies for task allocation in human-robot collaborative assembly. Following the descriptive analysis, a content analysis was conducted to answer the two research questions by categorizing the included references into two superior categories of approaches, namely static and dynamic, and investigating them regarding their methodologies for suitability assessment, task distribution, schedule optimization, and online adaptation, as well as their respectively applied criteria as a basis for task allocation decisions. The results help to understand the current state of the art in human-robot task allocation in assembly and give detailed insights into the specific methodologies proposed in literature.

Furthermore, based on the insights gained from the literature analysis, a reference model for task allocation was developed that encompasses the various categories of approaches identified. This reference model could form a basis for further research on this topic. Finally, current challenges that need to be investigated to expedite the implementation of task allocation methods in practice were discussed, and future research needs related to application scenarios, metrics and benchmarks, implementation barriers, acceptance, and complementary technologies were outlined.

However, despite the methodological procedure used in this review through a systematic literature search combined with a forward-backward search, there are some limitations of this study that need to be addressed in future reviews. First, although the selection of both search terms and databases was based on thorough initial research and the authors’ knowledge of the field, the selection and composition of search terms could exclude some relevant references, for example, due to additional terms used synonymously. In addition, an extension of the conducted study to other fields than assembly would be interesting and could reveal further approaches to task allocation that could also be transferred to the assembly sector.

In closing, task allocation is an important component of collaborative assembly and, if implemented effectively, has the potential to contribute to a significant increase in the use of human-robot collaboration in industrial assembly by enhancing productivity and reducing operator workload.

Disclosure statement

No potential conflict of interest was reported by the authors.