This paper aims to address the following main research questions: (i) What are the predominant research interests from the first cobot, i.e., 1996, to the end of 2022, and how can these be systematically classified?; (ii) which research topics in cobotics have garnered the most attention, and which studies and journals are considered most influential in this field?; (iii) who are the leading authors and which countries have made the most significant contributions to cobotics research?; (iv) how has the publication quantity in cobotics evolved over the years, and what trends can be predicted for future research output in this field?; and (v) what are the existing research gaps in cobotics, and what future research directions can be identified through a comprehensive scoping review and bibliometric analysis?

First, a comprehensive search strategy was formulated to access a diverse array of articles for consideration. This search was carried out in the ISI Web of Science (WoS), recognized as one of the premier repositories in research, covering the earliest date available in the database up to, and including, December 2022. The search string used was “collaborative robot*” OR cobot * in either the title, abstract, or author keywords fields. Due to the extensive nature of the research conducted, our objective was to identify as many relevant papers (including journal articles, reviews, and conference proceedings) as possible, focusing solely on cobots, totaling 2,195 records. That two-word search string was chosen after trying broader strategies leading to more records. Those broader strategies, including for example, the additional terms OR “human robot interaction” OR HRI multiplied the number of records by at least 40. However, a significant portion of those records was not related to cobots, as per its meaning presented in Section 1. Indeed, research on human-robot interaction or collaboration is also dealt with in the literature for industrial robots that are not inherently designed to interact with humans or for service robots for example. Adding other keywords made the amount of articles to screen impossible to assess humanly in a timely manner by the research assistants, due to limited staff, as well as time and budget constraints. Because of that, and to minimize the number of papers not related to cobots in an industrial context, the paper focuses only on the two-keyword strategy. Analyzing the papers started in 2023 and finished in 2024. The writing of the paper followed in 2025. Figure 1 illustrates the process of the PRISMA-ScR guideline for selecting the relevant literature and the number of papers handled at each part for the chosen strategy.

As can be seen in Figure 1, for conducting a scoping review, first, the focus was narrowed only to journal articles with abstracts, keywords, and DOIs, resulting in 857 documents. One record out of these 857 documents was excluded due to the reason of without full-text availability. Then, a screening process was used in order to ensure the exclusion of irrelevant articles. To minimize threats to validity, one team member conducted the screening process while another independently validated their work. Any disagreements during the screening were resolved through discussion between the two team members to reach a consensus and make a final decision. This screening process resulted in the exclusion of 323 papers, based on three specific exclusion criteria (see Figure 1): (i)

Beyond to cobot: 121 papers matched the search keywords but had a different definition from that provided in the introduction. Interestingly, among these articles, 20 focused on various types of robots (often aerial and/or terrestrial robots, and swarms of robots) working collaboratively, but without human.

After this screening process, 532 articles were validated for final analysis in this research.

In this paper, MS Excel and VOSviewer were used for the bibliometric analysis of the 532 retrieved papers. VOSviewer employs the VOS (Visualization of Similarities) mapping method to create networks where the distances between items indicate their degree of similarity. The VOS clustering technique categorizes topics into distinct clusters, each marked with a different color (van Eck and Waltman, 2010; van Eck and Waltman, 2007). These techniques allow for the analysis of papers in terms of co-authorship, co-occurrence of terms in the titles, abstracts, and keywords, as well as the bibliographic couplings of research institutions, countries, or publication sources. Thanks to these features, VOSviewer has been widely applied in the bibliometric analysis in various robotics fields (Wu et al., 2024; Wang J. et al., 2023; Mudhivarthi and Thakur, 2022; Long et al., 2024; Liu and Son, 2024; Li W-S. et al., 2021; Chu et al., 2021). Based on the key topics identified through the bibliometric analysis, a univocal categorization framework was proposed. Two team members reviewed the papers, classifying each to a category based on its main research objective. Any discrepancies were resolved through discussion.

The research topic categories were formed and organized by reviewing the papers and their research topics. Furthermore, MS Excel and VOSviewer were used again for bibliometric analysis on each category to gain more specific results for each research topics.

The number of publications is an important indicator for measuring the development trends within a research domain. By analyzing the number of publications over time, one can easily infer research activity levels and trends, which in turn allows for the estimation of future activity levels (Yang et al., 2019). Figure 2 presents the annual and cumulative number of publications focused on the topic of cobotics for industry. It shows that there were only 14 publications before or in 2015. This period can be considered the initial stage, where little research explicitly focused on cobot in industrial settings. Since 2016, the number of publications in this field has increased significantly. Furthermore, the cumulative number of publications approximately follows an exponential growth pattern, as calculated by Equation 1: C u m u l a t i v e n u m b e r o f p a p e r s n = a 1 × e a 2 n − 2000 (1) where n is the year for which we want to predict the cumulative number of papers, and a 1 and a 2 are coefficients estimated to be 0.02632 and 0.4514, respectively. Therefore, the cumulative number of papers in this field in a specific year can be calculated using Equation 2: C u m u l a t i v e n u m b e r o f p a p e r s n = 0.02632 × e 0.4514 n − 2000 (2)

The Coefficient of Determination ( R 2 ) is used to measure the fitting degree and can be calculated using Equation 3: R 2 = 1 − ∑ n = 1 N A c t u a l c u m u l a t i v e n u m b e r o f p a p e r s n − P r e d i c t e d c u m u l a t i v e n u m b e r o f p a p e r s n 2 ∑ n = 1 N A c t u a l c u m u l a t i v e n u m b e r o f p a p e r s n − M e a n o f a c t u a l c u m u l a t i v e n u m b e r o f p a p e r s n 2 (3) where N is the number of observations. An R 2 value closer to 1 indicates a better fit of the regression model, while a value closer to 0 indicates a poorer fit. Also, to measure the prediction error, the Root Mean Squared Error ( R M S E ) is used, as shown in Equation 4: R M S E = 1 n ∑ n = 1 N A c t u a l c u m u l a t i v e n u m b e r o f p a p e r s n − P r e d i c t e d c u m u l a t i v e n u m b e r o f p a p e r s n 2 (4)

The model demonstrates a high goodness-of-fit with an R 2 value of 0.995, indicating that the cumulative number of published papers in this field is growing exponentially, and that continued development in cobotics research is anticipated. The R M S E is 9.15, suggesting that the average prediction error is around 10 papers. The predicted cumulative number of publications is listed in Table 1. The number of publications is expected to reach approximately 2000 by 2025, and about 20000 by 2030.

The geographical span of a research field indicates its versatility and global popularity. The more institutions/countries involved, the more significant the research field. This aspect is explored in the following subsections.

The analysis of publication sources is essential to identify the main journals in a research field, to help researchers find relevant literature and select the right journals for their work. The analysis of 532 selected papers revealed that they came from 156 different journals. Table 4 lists journals that have published more than 10 papers on cobotics for industry. The journal “Robotics and Computer-Integrated Manufacturing” is the leading source, with 35 publications, followed by “IEEE Robotics and Automation Letters” with 33 papers, “Applied Sciences-Basel” with 29 papers, and the “International Journal of Advanced Manufacturing Technology” with 26 papers. Regarding citations, “Robotics and Computer-Integrated Manufacturing” ranks first with 1,602 citations, followed by the “International Journal of Advanced Manufacturing Technology” with 760 citations, and “IEEE Robotics and Automation Letters” with 734 citations. However, papers published in “Mechatronics” attracted the most attention on average, with 62.09 citations per paper, while “Robotics and Computer-Integrated Manufacturing” ranks second with 45.77 average citations. Figure 6 illustrates the bibliographic coupling among various publication sources active in the cobotics with at least 10 papers. The size of each sphere indicates the journal’s relative strength in publishing cobotics papers, the color represents clusters of journals, and the thickness of the arcs between journals indicates the strength of their bibliographic coupling.

Analyzing the number of publications and citations of authors, and mapping co-authorship relationship, allows for the identification of the most productive and influential authors and the most prominent research groups in the field, according to the PRISMA-ScR process applied for literature selection (see Figure 1). This information is invaluable for researchers seeking collaboration and insights from leading experts. The 532 retrieved papers were contributed by 1704 authors, with 26 authors having at least 5 publications, and 26 authors having a minimum of 200 citations (see Table 5). Understanding authors’ contributions to the cobotics research field requires evaluating both the number of articles and citations, and authors are ranked based on these metrics in Table 5. Even though this paper reviewed as many relevant papers as possible, the results are limited to documents retrieved through the PRISMA-ScR process described in Figure 1. For example, some publications may not be indexed in the WoS database, may have been missed due to the chosen search string, may have been excluded based on criteria such as being conference proceedings. These limitations could introduce inaccuracies. Another potential source of bias is the inability to distinguish between authors with identical names. Likewise, authors who publish under different names may not be correctly linked. To address this issue, it is strongly recommended that researchers be assigned a unique, persistent identifier, such as the Open Researcher and Contributor ID (ORCID), upon publishing their first paper. This would help resolve such issues, regardless of how a researcher’s name appears across different publications (van Nunen et al., 2018; Chiu and Ho, 2007).

Vidoni, R., and Vanderborght, B. lead the ranking in terms of the number of articles, with 8 papers each. Vidoni, a professor at the Free University of Bozen-Bolzano (Bolzano, Italy), specializes in high-performance (energy, vibration, safety, collaborative) automatic machines. Vanderborght, a professor at the Vrije Universiteit Brussel (Brussels, Belgium), focuses on cognitive and physical human-robot interaction, robot-assisted therapy, humanoids, and rehabilitation robotics using variable impedance actuators. In terms of citations, Secchi, C. leads with 600 citations. Secchi is a professor at the University of Modena and Reggio Emilia, known for his work in human-robot collaboration, multi-robot systems, and medical robotics.

Figure 7 illustrates co-authorship in the cobotics research field for industry, with spheres representing authors, colors indicating clusters of authors, and arcs showing co-authorship strength. It reveals that the international community actively engaged in cobotic research remains limited.

Analyzing the frequency of terms in the 532 retrieved papers offers valuable insights into the main research topics within the field of cobotics for industry. To identify the most prominent topics, a terms co-occurrence density map was constructed. Terms appearing in at least five papers were included, while general terms such as “work”, “model”, or “methodology”, as well as cobot-specific terms like “robot”, “collaboration”, “collaborative robot”, or “cobot”, were excluded. This process identified a total of 71 relevant terms, with their co-occurrence density visualized in Figure 8.

The category “Deployment of cobots” is how to deploy cobots in an industrial organization. Methodologies for deployment are studied to optimize cobots’ integration. Furthermore, comparing traditional robots versus cobots (Faccio et al., 2019), maximizing with other enabling technologies (e.g., virtual reality and digital twins) are some examples of research outputs. For instance, using a digital twin, which is a virtual replica (model) of a real system, can guide and enhance the performance of the actual system. Several researchers put efforts to explore the application of digital twins (Gallala et al., 2022; Lima et al., 2019; Malik and Brem, 2021; Ronzoni et al., 2021), virtual reality (Badia et al., 2022), mixed reality (Ostanin et al., 2021), or a combination of them (Pérez et al., 2020; Wolfartsberger et al., 2019), to address complexities in workspaces where humans and cobots interact with each other.

Growing interests on cobots in commercial markets has been investigated in the literature. These papers presented here explore the market potential and future prospects of cobots. For example, scholars have predicted that the significance of cobots will continue to grow (Bogue, 2022; Bogue, 2016b; Bloss, 2016). Although the cobotic is a relatively new research topic in the industrial robotics, the applications of cobots are rapidly expanding in industrial sectors such as assembly, packaging, and surface treatment. This category also covers the papers that use cases explained how to deploy efficiently cobots for studied applications. Table 6 presents an outline of the deployment of collaborative applications.

Despite increasing cobot-related applications, their adoption remains inadequate in real environments. Therefore, researchers investigate on effective parameters on cobots’ adoption such as trust between operator and cobot. Styles of human-robot symbiosis are modelled or experienced to improve trust and operators’ satisfaction as well as performance (Lambrechts et al., 2021; Zemlyak et al., 2022; Baumgartner et al., 2022; Quintana et al., 2022; Mateus et al., 2019; Andersson et al., 2021; Liu and Cao, 2022; Simoes et al., 2020; Bagheri et al., 2022; Kopp et al., 2022; Maurtua et al., 2017a; Sauer et al., 2021). Clearly enough, successfully deploying a cobot in a system depends on various factors and can have both positive and negative impacts. Therefore, evaluating the cobot’s impacts (e.g., economic, social, environmental impacts) before deployment is crucial. This category can contain works that investigate the impacts of cobots, as well as those that develop models at improving these impacts (Alvarez-de-los-Mozos et al., 2020; Ronzoni et al., 2021; Colim et al., 2021a; Palomba et al., 2021; Realyvásquez-Vargas et al., 2019; Chiriatti et al., 2022; Colim et al., 2021b; Zhang YJ. et al., 2021; Ojstersek et al., 2022; Calvo and Gil, 2022; Navas-Reascos et al., 2022; Javernik et al., 2022; Kim et al., 2021; El Makrini et al., 2022; Liu and Wang, 2020; Gualtieri et al., 2020; Ibáñez et al., 2021).

As discussed above, one of the advantages of implementing cobots is the reduction of health, ergonomic, and environmental risks (which is a social consideration), such as biomechanical overload. Industrial workers face to various ergonomic hazards such as improper postures, repetitive motions, heavy lifting, and vibrations. Several techniques are available to assess ergonomic risk factors, including observational, tool-based, and questionnaire-based methods (Cheraghi et al., 2019). Among these, observational methods like Rapid Upper Limb Assessment (RULA) (McAtamney and Corlett, 1993), Rapid Entire Body Assessment (REBA) (Hignett and McAtamney, 2000), Posture, Activity, Tools and Handling (PATH) (Buchholz et al., 1996), and the Ovako Working Posture Analysing System (OWAS) (Karhu et al., 1977) are widely used. These methods have been extensively applied for ergonomic evaluation of collaborative systems, including RULA (Colim et al., 2021a; Palomba et al., 2021; Chiriatti et al., 2022; Colim et al., 2021b; Navas-Reascos et al., 2022; Gualtieri et al., 2020; Ibáñez et al., 2021; Colim et al., 2021c), the Strain Index (SI) (Colim et al., 2021b; Zhang YJ. et al., 2021), NIOSH indices (Ronzoni et al., 2021), REBA (El Makrini et al., 2019), KIM-MHO (Colim et al., 2021b; Colim et al., 2020), EAWS (Maurice et al., 2019), JSI (Navas-Reascos et al., 2022), and OCRA (Ronzoni et al., 2021; Gualtieri et al., 2020). Despite their popularity, conventional techniques for evaluating ergonomic conditions in cobot-related workspaces have limitations: they may not cover all types of activities in a collaborative environment, and they are incapable of addressing dynamic phenomena, such as fast motions. Consequently, it is more accurate to employ a technique specifically designed for assessing ergonomic risks of collaborative jobs. To address this need, researchers can be attracted to propose methods specifically developed for evaluating ergonomic risks associated with collaborative activities (Maurice et al., 2017).

Grippers are the most widely adopted end-effectors in collaborative industrial applications. They are essential for tasks requiring precision and versatility like human-hand capabilities, making them an interesting research topic in the field of cobotics. Grippers must be designed to be cost-effective and capable of handling a wide range of objects that vary in size, shape, and stiffness. In addition, they should ensure operator safety by limiting contact forces and avoiding sharp edges and pinch points. This sub-category has been attracted a lot of attentions of the cobotic researchers (Bogue, 2016a; Birglen, 2019; Iqbal et al., 2021; Anwar et al., 2019).

This category represents how to guarantee the safety of the operator while optimizing the performance of the cobot. It is mainly dedicated to the risk of an unwanted collision between the cobot and the operator. This category comprises papers discussing standards for cobot safety and how to meet them. According to those papers, the main safety standards related to robotic are Section IV of Chapter 4 of OSHA Technical Manual (OSHA, 2017), ANSI/RIA R15.06 (Association, 2012), ISO 10218-1 (ISO, 2011a), and ISO 10218-2 (ISO, 2011b). Among these standards, the most relevant standard addressing cobots in addition to conventional industrial robots is ISO 10218, Parts 1 and 2. While the papers analyzed regarding the safety category used the 2011 version of ISO 10218, a recent version of that standard was issued in 2025, i.e., after the publication of those papers. The 2011 version of part 2 introduced the concept of “collaborative robots”, which is no longer used in its 2025 version (ISO, 2025a; ISO, 2025b). The latter focuses more on the collaborative aspect of the robot application. However, to design a safe collaborative workspace, a technical specification ISO/TS 15066 (ISO, 2016), that complements the former version of ISO 10218 was needed. Now, most of that TS is available and updated in the 2025 version of ISO 10218.

The ISO 10218:2011 and ISO/TS 15066:2016 cited in the papers introduced four safety methods for a collaborative application: (i) safety-rated monitored stop (SRMS), (ii) hand guiding (HG), (iii) SSM, and (iv) PFL. The 2025 version of ISO 10218 now only considers methods (ii) to (iv) for collaborative applications. Even though SRMS is not in the 2025 version, the following paragraph describes succinctly each of the four methods for better understanding of their meaning.

In the SRMS method, both the operator and cobot could not operate simultaneously within the shared space. To avoid an unwanted collision, the cobot would have been stopped when it detects an operator inside the shared space. In the HG method, the operator directly teaches the cobot by physically moving it. In the SSM method, the cobot’s speed adjusts based on the separation distance between the operator and the cobot. The PFL method is prone to unintentional collisions. It limits the quasi-static contact (i.e., contact between an operator and part of a robot system, where the operator body part can be clamped between a moving part of a robot system and another fixed or moving part of the robot cell) and transient contact (i.e., contact between an operator and part of a robot system, where the operator body part is not clamped and can recoil or retract from the moving part of the robot system) forces below specific thresholds to mitigate human injuries. Achieving that involves active safety measures like compliance control and passive safety measures like edge smoothing during the cobot’s design stage. Based on ISO 10218-2:2025, hazard identification and risk analysis is needed to identify the hazards and assess the risks of both the cobot system and the collaborative environment and then select the proper safety measures (ISO, 2011a).

The developed methodologies in the literature related to the risk of collision between a cobot and an operator can be classified into two main groups: (i) pre-collision (collision prevention), and (ii) post-collision (collision mitigation). The pre-collision approach aims to ensure the safety of operators by detecting potential collisions between the cobot and the operator. This detection is achieved through safety sensor systems such as proximity and visual sensors, or by predicting unintended collisions using stochastic or machine learning methods. Once a potential collision is identified, preventive measures are implemented to avoid it (Li et al., 2023; Gualtieri et al., 2021a; Zorman et al., 2023). In the context of pre-collision strategies, the former SRMS, as well as the actual HG, and SSM safety methods played a key role in the papers analyzed.

Dynamic monitoring allows to know where are located the robot and the operator in unstructured and time-varying environment using advanced sensing technologies. Various approaches have been developed, including the use of virtual reality, Internet of Things (IoT), and sensing technologies to monitor humans in robot workspaces, aiming to prevent collisions (Tsuji and Kohama, 2019; Farsoni et al., 2022; Guerra et al., 2016; Gradolewski et al., 2020; Geiger and Waldschmidt, 2019; Ferraguti et al., 2020; Costanzo et al., 2022; Bin Islam et al., 2019; Barbosa et al., 2022; Safeea and Neto, 2019; Benli et al., 2019; Kianoush et al., 2021). Effective collision avoidance and path optimization are crucial for maintaining safety and productivity in collaborative applications. Therefore, several studies have focused on pre-collision algorithms and dynamic path planning to minimize collision risks and improve cycle times (Scimmi et al., 2021; Kot et al., 2022; Scalera et al., 2022; Chiriatti et al., 2021; Zanchettin et al., 2022; Chan and Tsai, 2020). In addition, the visibility of the robot to the human operator can significantly impact both the operator’s comfort and the probability of collisions. This critical factor has been examined in several research papers with the aim of enhancing the safety of collaborative application (Tarbouriech and Suleiman, 2020; Dufour et al., 2020; Najmaei et al., 2010). Readers interested in further exploration of collision avoidance can find successful efforts in (Simas et al., 2022; Scoccia et al., 2021; Safeea et al., 2019; Elguea-Aguinaco et al., 2022).

On the contrary, the post-collision approach focuses on operator safety through the detection of unintended collisions and then mitigation of energy transfer during these collisions that can lead to human injuries or even fatalities (Li et al., 2023; Gualtieri et al., 2021a; Zorman et al., 2023). In the context of post-collision strategies, the PFL method is employed. As mentioned above, although the PFL method allows unintentional collisions, it is required to include contact force measurements to validate safety (Zimmermann et al., 2022). Therefore, techniques (such as statistical models) to determine contact force thresholds for using in PFL method has been proposed (Behrens et al., 2022). Additionally, several testing procedures for validating safety in PFL method have been developed (DGUV, 2017; ANSI, 2018), and the reliability of such tests have been examined (Scibilia et al., 2021).

Soft cobots (i.e., cobots that can deform and yield in a collision) are often equipped with a protective covering, known as “skin”, to offer both active protection (detecting collisions and initiating cobot reactions) and passive protection (providing a cushioning effect to mitigate the consequences of collisions). Therefore, developing cobot skins take a great apportion of studies in this category (Runge and Raatz, 2017; Ye et al., 2022; Svarny et al., 2022; Pang et al., 2021; Nguyen et al., 2021; Ge et al., 2022; Ye ZQ. et al., 2020; O'Neill et al., 2018; Tsuji and Kohama, 2020; Heng et al., 2021). Furthermore, to reduce cobot production costs, sensor-less collision detection methods (e.g., neural network-based approaches) have gained attention (Kwon et al., 2021; Lee and Song, 2016; Heo et al., 2019; Sharkawy et al., 2020a; Ma et al., 2020; Czubenko and Kowalczuk, 2021; Le and Kang, 2022; Yen et al., 2019; Kim, 2022). Many other studies have contributed to the field of collision detection, including (Heo et al., 2019; Abu Al-Haija and Al-Saraireh, 2022; Amin et al., 2020; Huang et al., 2022; Lee et al., 2019; Li W. et al., 2020; Long et al., 2022; Lu SN. et al., 2022; Nguyen and Case, 2022; Park et al., 2022; Ren et al., 2018; Rodrigues et al., 2023; Shin et al., 2019; Strazdas et al., 2021; Xiao et al., 2018; Yun A. et al., 2022; Zhang ZJ. et al., 2021), providing valuable information for researchers. To ensure the operator’s safety in a collaborative workspace, in addition to measuring and controlling the interaction force, other parameters like its direction, the point of application, and the time required to measure these parameters are also important and can form the research objective of a study (Popov et al., 2021).

Several techniques have been developed for identifying and evaluating system hazards, such as Failure Mode and Effect Analysis (FMEA), Job Safety Analysis (JSA), and Fault tree analysis (FTA), as well as for assessing human reliability, including Human Error Assessment and Reduction Technique (HEART) and Cognitive Reliability and Error Analysis Method (CREAM). These generic techniques can be applied to cobotics. However, due to the unique challenges inherent in cobotic applications, some researchers have worked on developing safety metrics and risk assessment tools specifically for these systems (Askarpour et al., 2019; Marvel et al., 2015; Pantano et al., 2022; Vemula et al., 2018; Vicentini et al., 2020). A safety metric or risk assessment tool for cobotics should be able to consider various parameters, such as the uncertainty associated with human operator behaviors, the potential severity of physical contact between the cobot and the operator, and changes within the system. Last but not least, one of the key factors that can impact safe collaboration in cobotics is the threat of cyber-attacks. Defense strategies to protect workers against cyber-attacks have been also investigated (Khalid et al., 2018).

“Human-robot tasks allocation” aims to distribute the tasks between humans and cobots. This involves determining the sequence of tasks and deciding which agent (cobot or human) should perform each task. With shorter product life cycles and high-mix production, the job splitting between the agents is becoming increasingly important. Research in this field has focused on optimizing task allocation to enhance efficiency, ergonomics, and economic factors. Optimization models aim to minimize makespan and production costs while considering ergonomic risks and worker wellbeing (El Makrini et al., 2019; Yu et al., 2020; Wang et al., 2018; Nourmohammadi et al., 2022; Mura and Dini, 2022; Liu et al., 2024b; Liau and Ryu, 2022a; Gjeldum et al., 2022; Cai et al., 2023; Ferreira et al., 2021; Kinast et al., 2022; Almasarwah et al., 2022; Stecke and Mokhtarzadeh, 2022; Mura and Dini, 2019; Li ZX. et al., 2021; Weckenborg et al., 2022; Pearce et al., 2018; Boschetti et al., 2021a).

As revealed in the above-reviewed papers, optimal human-robot task allocation involves balancing objectives like minimizing cycle time, costs, and ergonomic risks. Selecting the best approach is challenging due to conflicting requirements. For instance, minimizing production costs by assigning more tasks to cheaper work cells may delay deliveries, while minimizing makespan may increase costs. Additionally, relying solely on these objectives may lead to assigning high-risk tasks to human workers. Thus, the collaborative applications designers should comprehensively consider economic, social, and environmental aspects to ensure sustainable task allocation in collaborative applications.

These approaches are powerful in a deterministic collaborative environment, where both the operator and the cobot must adhere to a pre-planned task sequence. However, in real collaborative task-allocation problems, various uncertain parameters, especially those related to the operator, should be considered. For example, the operator may deviate from the task sequence, either by mistake or based on personal preference. Such deviations may not yield the predefined optimal solution but could have a negligible impact on the objectives of the system. In such scenarios, the process may halt if the cobot cannot adapt to new conditions or human actions. Therefore, a cobot should be capable of finding alternative solutions while still maintaining the original objectives. To address these uncertainties, dynamic task allocation models have been developed (Antonelli et al., 2021; Pupa et al., 2021; Petzoldt et al., 2022; Chacón et al., 2021; Messeri et al., 2022; Casalino et al., 2021; Antonelli and Bruno, 2019; Bruno and Antonelli, 2018; Pupa et al., 2022).

“Human-robot interaction” gathers research on the way that the human and the robot interact themselves (co-working) in order to perform their allocated industrial actions. In the collaborative application, the operator often defines the objective of collaborative work and the cobot assists the operator to achieve this objective. The cobot must therefore be able to learn from the operator and estimate his intentions. Various input interaction modes are explored to instruct the cobot about the state of humans or their intentions such as gesture visual recognition and force sensing.

Gesture (e.g., hand, body, and head) visual recognition is one of the most powerful communication mode which can be developed by techniques such as wearable sensors, inertial measurement unit (IMU) sensors surface electromyography (EMG) signals (Mukherjee et al., 2022; Papanagiotou et al., 2021; El Aswad et al., 2021; Coupeté et al., 2019; Mendes, 2022; Cornak et al., 2022; Tang et al., 2015; Shukla et al., 2018; Lee et al., 2022). Besides vision-based systems, wearable systems capable of measuring joint rotation can easily capture an operator’s gestures. However, a primary limitation of such systems is their potential interference with the operator’s work. In addressing this challenge, wireless wearable system designed to determine the orientation of the operator’s upper body parts has been developed (Skulj et al., 2021). Tactile sensing allows cobots to detect and interpret pressure and force exerted by human operators (Olivares-Alarcos et al., 2019; Bauer et al., 2008). Cobotics leveraging tactile sensors can dynamically determine human grasp positions and intentions being able to ability to adapt to human actions and improve overall interaction efficiency (Li TJ. et al., 2021; Ansari et al., 2020). Overall, numerous papers in this category focus on designing and developing new sensors to improve cobotic systems and increasing their cost-effectiveness (Zaid et al., 2022; Long et al., 2021; Fu and Cai, 2022; Castano-Cano et al., 2022). In human-robot interactions, relying solely on mechanical force-torque sensors to measure forces can sometimes lead to the unintentional measurement of additional forces arising from the robot’s contact with an unpredictable environment. This can potentially introduce inconsistencies in human-robot interactions (Ajoudani et al., 2018). As an alternative approach, bio-signals, like EMG signals, can be used to directly measure the forces exerted by the human operator (Bednarczyk et al., 2022).

In a collaborative application, communication methods extend beyond gesture recognition and force sensing to include natural speech, gaze, graphical signage, and physiological signals (e.g., electroencephalographic (EEG)) (Dmytriyev et al., 2022; Eimontaite et al., 2022; Eimontaite et al., 2019). Natural speech and gaze are also crucial for effective human-robot interaction. Speech control systems allow operators to command multiple robots in different languages (Hofer and Strohmeier, 2019). Also, using EEG signals, emotional states (e.g., fearful) of operators can be assessed in a cobotic system (Buerkle et al., 2021; Borboni et al., 2022; Eyam et al., 2021). In addition, to enhance communication reliability, some researchers have adopted multimodal interaction modes to provide complementary or redundant input options (Tsuji and Kohama, 2022).

“Performance of actuating systems” is how to improve the actuating system of a cobot through its logical or physical equipment. The control of trajectory and motion is addressed in this category. To achieve high efficiency, precision, and safety in cobotics, numerous researchers have made efforts to develop trajectory and motion planning methods. These methods aim to balance performance with psychological and physical safety of the operator (Zhao et al., 2022; Ye L. et al., 2020; Wittmann and Rixen, 2022; Vysocky et al., 2020a; Sidobre and Desornneaux, 2019; Magyar et al., 2019; Li et al., 2019; Krämer and Bertram, 2022; Kraemer et al., 2020; Chen et al., 2018; Liu et al., 2022a; Rojas et al., 2019; Rojas et al., 2020; Zhu et al., 2022; Palleschi et al., 2021).

Force control was originally developed for applications where the robot’s end-effector comes into direct physical contact with its environment, often involving tasks like surface treatments, such as polishing and grinding (Ochoa and Cortesao, 2022; Ubeda et al., 2021; Gracia et al., 2019; Lakshminarayanan et al., 2021; Perez-Vidal et al., 2019). These scenarios require precise force management to achieve the desired quality of work. Due to the safety concerns, force control and its variants, such as impedance control and admittance control, become increasingly significant in a collaborative system. Impedance control involves computing the robot’s resistance to motion when external forces are applied, while admittance (i.e., the inverse of the impedance) control focuses on adjusting the robot’s trajectory in response to external forces (Siciliano and Villani, 1999; Khan et al., 2014). Impedance/admittance control is commonly employed in cobotics for physical interactions among humans, cobots, and the environment (Aydin et al., 2021; Li et al., 2020c; Kim and Yang, 2021; Dou et al., 2022). Compliance (the opposite of stiffness), i.e., the ability to yield or adapt to external forces, can be achieved through (i) active compliance, and (ii) passive compliance. Active compliance is a software-based approach, allowing the robot to dynamically respond to external forces. Passive compliance, on the other hand, is designed into the robot’s mechanical structure, allowing it to naturally respond to external forces. Both approaches have been extensively investigated in the literature (Huang et al., 2019; Huang et al., 2020; Huang et al., 2021; Zhang SL. et al., 2021; Zeng F. et al., 2019).

The ability to vary stiffness is crucial in cobotics. High stiffness is employed for regular operational routines, while reduced stiffness can minor contact forces or even be a detector of collisions (Ayoubi et al., 2019; Ayoubi et al., 2020; Liu YW. et al., 2021). In addition, variable stiffness enables the cobot to perform diverse tasks (Ren et al., 2019a). Clearly enough, accuracy is a critical performance characteristic of a cobotic system, particularly in applications that demand precise positioning, manipulation, or sensing. Calibration is the process that guarantees the cobotic system achieves the desired accuracy. Research in this area focuses on improving metrology system and calibration techniques (Pagani et al., 2021; Yun HT. et al., 2022; Kim et al., 2022; Pollák et al., 2020; Kolyubin et al., 2019).

In the literature, some concepts of novel architecture of cobots are explored, such as modular components between cobots, two armed cobot, or extra components to reduce inertia. For instance, the integration of two cobotic arms significantly enhances the potential of cobotics for industrial applications, allowing for tasks such as handling bulky and heavy objects (Liu LY. et al., 2021; Cherubini et al., 2019). In addition, as discussed above, one of the most significant advantages of cobots over traditional industrial robots is their adaptability to a wide range of tasks. However, achieving this flexibility requires more than just reprogramming; the hardware components also need to be reconfigured (Romiti et al., 2022; Smrcek et al., 2022; Schou and Madsen, 2017; Wojtynek et al., 2019).

“Robot program generation” is how to generate the program for the cobot, by benefiting of the collaboration with humans. It is useful to enable a fast and easy-to-attain cobot task reconfiguration, even for inexperienced operators. Programming by demonstration develops a manual physical guiding of the cobot by its end-effector. System records waypoints, force inside joints or gripper. This method has been widely applied in the literature (Canfield et al., 2021; Ochoa and Cortesao, 2022; Wang et al., 2019a; Wang LK. et al., 2022; Wang LK. et al., 2023; Tadese et al., 2021; Svejda et al., 2022; Stenmark et al., 2018; Rozo et al., 2016; Müller et al., 2020; Iturrate et al., 2021; De Winter et al., 2019; Carfi et al., 2020; Al-Yacoub et al., 2021a; Ajaykumar et al., 2021; Wang et al., 2021; Soares et al., 2021; El et al., 2022; Halim et al., 2022; Safeea and Neto, 2022). While “programming by demonstration” requires a lower level of knowledge from the operator, a cobot’s functionality is typically limited to replaying recorded actions. To overcome this limitation, teaching the cobot from visual input through demonstrations has been developed in the literature (De Coninck et al., 2020).

Skill-based programming reaches the same objective of quick and easy programming, enabling cobots to handle a wide range of tasks. In skill-based programming, robots follow a hierarchy that involves lower-level entities called “robot skills”. These skills represent specific actions that the robot is capable of executing. A task (e.g., removing a specific object from a table) is a sequence of the skills (e.g., pick up object) along with specified parameters (e.g., the object to pick up). Skills are pre-defined by the robotics programmer. The sequence of skills and their associated parameters, required to complete a task, is explicitly programmed by non-expert robot operators in the factory (Schou et al., 2018; Pedersen et al., 2016). Skill-based programming has gained the attention of several research studies (Giberti et al., 2022; Polverini et al., 2019; Herrero et al., 2017).

Each of the 532 retrieved papers was classified according to its main research objective. Figure 10 displays both the annual and cumulative number of publications for each research topic. The trends observed conclude that the exponential growth previously identified in Figure 2 is consistent across all research topics. It is evident that publications across all research topics have become more regular since 2016. Also, Table 7 provides the number of papers for each research topic, along with citation analysis and details on the most productive authors, institutions, countries, and leading journals. The most popular research topics are “safety in cobotics for industry” with 151 papers, followed by “deployment of cobots” with 141 papers and “performance of actuating systems” with 102 papers.

Being aware that this paper presents an analysis of the scientific literature up to the end of 2022 in a fast-growing field, due to the constraints explained in Section 2.2, we present in this section some references, that give an idea of the trends in that domain for 2023 and 2024.

The International Federation of Robotics (IFR) (IFR, 2024) stated the top five global robotics trends in 2024: 1) AI and machine learning, 2) cobots in new applications, 3) mobile manipulators, 4) digital twins, and 5) humanoids. While trend 2 is about cobotics expanding in areas such as welding, trends 1, 3, and 4 can serve collaborative applications. Indeed, according to IFR, further cognitive collaboration with humans will be possible thanks to AI. Semeraro et al. (Semeraro et al., 2023) agree that machine learning has a great potential in HRI since it is a new way to develop cognitive models and behavioural blocks. IFR thinks that the possibility of mounting cobots onto mobile platforms generates new opportunities that will increase the demand for cobots. As shown in (Zafar et al., 2024), digital twins can serve collaborative applications by making them safer, through testing and optimizing the approach for HRI thanks to real-time data.