1 Introduction

Worker and industry perception of robots can greatly influence their adoption in practice. For this reason, the actual appearance and presentation of robots is important. What constitutes good robot design? According to a study by DiSalvo et al. (2002), the presence of facial features greatly influences the human perception, or first impression, of robots. The study further elaborated the threshold in which perception is positive and when it is “uncanny”, or too human. General user principles, such as the Technology Acceptance Model (TAM, Davis 1989) aim at mitigating the risk of technology rejection and abandonment by the user (de Graaf et al. 2017). Some more recent studies in social sciences and health care robotics have further investigated the design attributes with a focus on the user perception of the robots’ visual attributes (Reeves et al. 2020; Klueber and Onnasch 2022).

Concurrently, there is a strongly increasing trend of robotic applications in the construction sector. In 2021, a leading producer of robotic equipment and automation technologies conducted an international survey with 1900 small, medium and large construction and associated businesses, such as planners (architects, designers), component suppliers, contractor, subcontractors and consultants (ABB Robotics 2021 Construction Survey). The majority (91%) of respondents predicted a skills crisis in the construction sector by 2030. 81% of the respondents (or more than 1500 companies) will introduce robotic systems in the next 10 years, out of which 65% stated that they will introduce on-site robotics, compared to only 20% for off-site fabrication and merely 15% for 3D-printing. Participants identified concerns on health and safety, and environmental distress as catalysts for accelerated investments into construction robotics.

Current construction robots are commonly classified as single-task robots, general purpose robots and integrated automated robotic systems (Pan and Pan 2020; Garcia de Soto and Skibniewski 2020). Due to the limited capacity (e.g. autonomy, payload, navigation) of today’s robotic systems, Human–Robot Systems (HRS, e.g. collaborative robots) can play an important role for the low to medium size batch production predominant in AEC. Multiple designs have been proposed over the years, some of which aim to combine performance and human safety (Zinn et al. 2004). The level of interaction and collaboration varies according to HRS technology and application (Hentout et al. 2019). Recent academic studies by Kyjanek et al. (2019) and Mitterberger et al. (2022) investigate interactive fabrication process application for HRS in AEC that include the use of augmented reality interfaces for industrial robots in on-site construction and prefabrication settings, such as an off-site factory floor. Brosque and Fischer (2022) demonstrated that construction robots can have a positive impact on human safety. Their study further indicates that project quality, schedule and costs can be positively influenced by the introduction of construction robotics.

However, psychological perspectives on robotics and automation in construction have not gained as much attention in AEC scholarship when compared to research focusing on technology development and application engineering. Today’s construction field is still a highly manual work industry and it is not clear how the current human workforce will use, accept or adopt new robotic technology (Muishout et al. 2020). More particularly, construction field workers are prone to risk-taking behaviour (Low et al. 2019) and reluctant towards vocational skill training (Wang et al. 2018). Currently, we lack an understanding of individual AEC stakeholder groups’ perception of HRS. For greater HRS adoption in construction, one fundamental question is to understand how user acceptance of and engagement with robots may be impacted by visual and morphological aspects of the system’s design.

Therefore, this empirical study seeks to identify relationships between objective, visual attributes and individual, subjective perceptions of construction robots. The insights may suggest best practices for existing and emerging construction robotic design considerations to help increase the likelihood of positive AEC user perception, acceptance and ultimately adoption using participatory design principles (Muller et al. 1993), e.g. for and with construction worker safety in mind (Gambatese et al. 2005). Additionally, principles of work design and hedonic motivation should be considered as worker’s “safety” should not only focus on the avoidance of physical danger.

2 Theoretical point of departure

Muishout et al. (2020) conducted exploratory interviews with ten stakeholders (research academics, 3D-concrete-printing companies and labour union staff) in the Dutch AEC sector. The authors found that there is still an unclear general definition of the term “robotisation” and related socio-economic factors amongst these stakeholders. According to a study with 63 construction project managers by Sam et al. (2022), respondents were more likely to consider a robot as a teammate, as opposed to a machine, tool or equipment when the robot moved in a predictable way, was considered durable, remained constantly active, took its surroundings into account before moving, simply worked well when alongside human workers, was reliable and made the particular task more predictable.

To further identify trends in robotic construction scholarship across several decades (1988–2021), we manually collected 59 related academic articles (see supplementary materials) in the field of AEC for a time-trend analysis. We added robot* and construction to the list of stop words. The results here are two-fold: The term human has not been in the Top 10 most frequent words in most periods, but ranked 2nd in 2020 and 1st in 2021 (Table 1). Second, we are able to find correlations of keywords. For example, the term human showcases strong bibliometric correlation with the terms characteristic, sequence, intelligence and sensor (0.99), hand and performance (0.97), actions (0.94) and services (0.90). Surprisingly, however, our query shows that the term design was only placed 8th before 2000 and 2008–2014, 7th in 2018 and is not present in any of the more recent articles. This indicates that even though human has become a more frequently used term in our corpus, the term design has eventually dropped out. The terms perception, impression or ethic* could not be found in any of the most frequent words in this data set but have been discussed in scholarship outside of AEC.

Table 1 Top 3+2 word frequencies in the selected corpus
Full size table

For example, Knight (2014) promoted five partnership principles of robotic design, namely to: (1) Implement safeguards to ensure robots augment human experience rather than increase social barriers. (2) Account for the consequences of a robot’s social abilities when designing industrial guidelines and technologies. (3) Consider regulating the machine’s treatment as well as encouraging positive experiences for the user. (4) Design some robots to be intentionally machine-like since humans view robots as agents and react to them socially. (5) Decide whether a more machine-like or human-like design is most beneficial for the human context.

As construction robotics enter the growth phase on their way to ubiquity (Bock 2015), it is important to discuss the social-economic impact of those technologies on the workforce. Recent estimates state that roughly 50% of all construction tasks can be automated. This would indicate a replacement or displacement of 2.7 million construction workers in the US alone until 2057, and a reduced labor income by a net of US$31.5 billion (Manzo et al. 2018). At last, automation and new technologies in construction will challenge many existing roles and create new ones, eventually attracting an emerging generation of tech-savvy workers to AEC (Garcia de Soto et al. 2019). Technologies have been transforming work throughout the economy for many decades (Badham 2009). In some cases, technology can assist workers in performing jobs faster or more efficiently. In other cases, however, some human workers are simply being replaced by machines. According to Rumberger (1984), job opportunities are put at risk by a combination of forces, out of which high technology and automation are the most threatening. Furthermore, poor physical health and mental distress are predicting indicators of long-term unemployment in construction worker cohorts (Leino-Arjas et al. 1999), perpetually increasing the worker’s stress levels and impacting the sector’s productivity negatively.

3 Problem statement and research gap

Today, users may encounter cartoon-like, googly eyes placed on field robots, including in the context of construction (Fig. 1) .

Fig. 1
figure 1

Googly eyes on Dusty Robotics Prototype (2020)

Full size image

Although this simple strategy proves to be even successful with inanimate objects (Powell et al. 2012), it indicates a limited understanding of user-driven design attributes. Scholars such as van de Poel (2013a) have widely discussed that new technology development should be considered a social experiment, not an engineering task. Therefore, it is paramount to understand that when social and psychological considerations are not addressed, human workers might refuse to collaborate with or abandon such systems in the long term (e.g. de Graaf et al. 2017).

Given the delicate position of the construction industry facing automation and the related fear of job loss, a better understanding of barriers to the introduction of robots (incl. HRS) for the construction field should be considered important and urgent. Both studies by Muishout et al. (2020) and Sam et al. (2022) provide valuable insights into the human systems engineering (HSE) subject of robots in construction. The main restrictions, however, are the limited numbers of stakeholder groups (only four) and robotic systems under study (only three). Looking beyond AEC scholarship, Klueber and Onnasch (2022) have conducted a study to research the perceived “likeability” of feature combinations for social care robots in healthcare. The image-based online survey design combined and compared appearance attributes (human-like, animal-like, machine-like), size (small, large) and communication (speech, sound, text) using best-worst scaling. So far, to the best of knowledge, no similar study has been conducted on the appearance of construction robots with construction professionals.

4 Research question and hypothesis

Accordingly, we pose the research question: How do visual attributes influence AEC stakeholders’ perception of construction robots? Further, we address the need to understand key attributes and key perceptions of particular construction robots from various viewpoints. Hence, we formulate our main hypothesis: “The attributes morphology, color and material have an impact on the defined perceptions of ease of use, work task adaptability and risk of job loss in various stakeholder groups within the AEC industry.”

5 Methods

The study by Reeves et al. (2020) studied social robot attributes and perceptions in two separate studies that are based on a self-developed image database. Similarly, our methodology consists of a research approach involving the following steps: (1) Qualitative data collection via semi-structured interviews, (2) Definition of attributes and perceptions, (3) Development of image database, (4) Attribute validation study of defined attributes, (5) Quantitative perception survey with defined perceptions, (6) Statistical data processing and (7) Presentation of results and interpretation (Fig. 2).

Fig. 2
figure 2

Flowchart of the applied research method

Full size image

(1) Interviews First, we conducted five qualitative, semi-structured interviews with five experts in different sub-areas of AEC (architects, professors, product- and branch managers, researchers and engineers). The interview duration ranged between 30 and 45 min. Recruiting of the participants was done by convenience and proximity in the research network on a voluntary basis and without monetary compensation. The interviews were conducted in German and English and transcribed verbatim leading to 21 pages of interviews. The interview questions are available in Appendix 1. Table 2 shows the evaluated themes that were collected during the interviews. Productivity and price/cost—all economic viability terms—are not considered due to them not primarily being visual objectively measurable.

Table 2 Qualitative evaluation interviews themes (N = 5)
Full size table

(2) Based on the interviews, key characteristics are evaluated qualitatively: Objective features are referred to as “attributes”, whereas subjective features are referred to as “perceptions”. Table 3 illustrates the chosen set of three objective attributes and subjective perception chosen due to a combination of interview results and literature review. Our final research methodology considers three attributes with each two to three expressions, such as in the study design of Klueber and Onnasch (2022): (a) Form factor, or Morphology: anthropomorphic (human-like), zoomorphic (animal-like), mechanomorphic (machine-like). (b) Color: bright, dark. (c) Material: soft, hard. Perceptions: (d) Ease of use (inspired by Davis 1989). (e) Work task adaptability (or production flexibility, such as studied by Hartmann et al. 2009) and (f) Risk of job loss (Brand 2015).

Table 3 Selected attributes and perceptions
Full size table

(3) Image database Inspired by the Stanford Social Robot Database (2020) we developed a database of 50 construction robotsFootnote 1; 86% are existing machines, 14% are virtual visualisations and do not exist (yet). Surveys that rely on images are easy to understand, more engaging and thus tend to lead to higher response rates (Treiblmaier and Filzmoser 2011). The images in the database are subjectively chosen by the authors to achieve a balance according to the three defined attributes and expressions, e.g.: for each robot of attribute 1, expression 1 there shall be one of attribute 1, expression 2. The images are prepared in a resolution of 400x400px. Human operators, other background objects, brands and the environment have been removed to standardise the visual stimuli and thereby stimulate a spontaneous and mostly unbiased perception. Figure 3 shows a selection of the database entries. The full data set can be found in Appendix 3.

Fig. 3
figure 3

Excerpt of the developed robot picture database (n = 50)

Full size image

(4) Attribute validation study In this first survey, the robots’ visual characteristics are measured. We developed a quantitative online study on the survey platform Qualtrics\(^{\mathrm{TM}}\) that collected data on the attributes of 20 randomly chosen pictures of the 50 robots in the database. No relation to the broader AEC sector was required for the participants and recruitment occurred by convenience within private networks. This survey used the following items for validation by the participants: Morphology (anthropomorphic, zoomorphic, mechanomorphic), color (bright, dark), material (soft, hard) and was conducted during 4 weeks in Spring 2022. The survey was available in German and English with an estimated duration of 15–20 min. A cut-off value of 0.8 is used to determine if an attribute can be treated as objective by the majority of participants (e.g. see Gaul et al. 2010): Attribute expressions with less than 80% agreement between participants were considered unclear in our data set and were not further used for analysis.

(5) Quantitative perception survey For an eight week period in Spring 2022, we conducted a quantitative study on the same platform. For this survey, however, participants from the AEC industry were recruited via direct email outreach to more than 200 Swiss construction companies, physical posters at construction sites and direct messages within the authors’ extended professional, social and research networks. The survey was designed to be intuitive and interactive for the participants and responses were collected via a 5-point Likert scale. Participants were shown a random sample of 20 of the 50 robot images in the data set to reach an assumed completion time of 15–20 min. They were asked to judge three predefined perceptions in the context of the construction industry: (a) Ease of use: How easy to use does the robot appear to be? (b) Work task adaptability: How flexible is the application of the robot? In particular, we asked if the machine appears to have only a single or several use cases. (c) Risk of job loss: How likely is the robot to replace humans?

(6) Statistical data processing Binary logistic regression (BLR) is used to estimate the relationship between one or more independent variables and a binary dependent variable. BLR is chosen to be appropriate to prove or disprove our binary outcome hypothesis as probabilities are bounded by 0 and 1. In our approach, both independent and dependent variables are dichotomous (mutually exclusive, collectively exhaustive). By evaluating Exp(b), or the Odd Ratio, conclusions can be drawn on the relationship between attributes and perceptions which are the main interest of this study. Further information on BLR can be found in Appendix 2. Raw survey data are exported from the Qualtrics\(^{\mathrm{TM}}\) platform, imported into IBM\(^{\mathrm{TM}}\) SPSS Statistics Software and prepared for BLR analysis. Systematic errors are evaluated by scanning the data for systematic response bias. This is done by evaluating the standard deviation from every recorded response. If the standard deviation SD = 0, the response is considered invalid since the participant always chose the same answer for the entire survey. Incomplete responses of the designed online survey platform are considered “completely at random”. According to Leung and Qin (2006) the missingness of data under completely random state can be ignored and, therefore, are kept in the data set. To compare the different characteristics with BLR, every attribute type is coded with a binary parameter: Consequently, the 5-point Likert scale of the quantitative perception survey is compressed into net positive and net negative binary values while neutral values are ignored.

6 Results

Table 4 shows the number of expressions that were collected by the attribute validation study (N = 142) with a cut-off value of 0.8. Note that most robots are considered of mechanomorphic morphology and only one robot is considered soft. This emphasizes the importance of the validation of the stimulus material for experimental settings.

Table 4 Results of the attribute validation study (N = 142)
Full size table

Table 5 shows the results of the attribute validation study for ten (out of 50) prominent examples of construction robots and common equipment. The full table with the results of the attribute validation study and quantitative perception survey can be found in Appendix 3.

Table 5 Descriptive statistic data from the attribute validation study
Full size table

Quantitative perception survey: In total, 251 participants of 12 different stakeholder groups participated in the subjective perception survey. Table 6 shows the professional distribution of the respondents, some of which reported more than one profession. 60 participants never completed the entire survey, 25 did not specify a stakeholder group and 5 created an invalid systematic error. Thus, a total of 161 valid responses are collected and analyzed. The respondents’ demographics are presented in Table 7. The majority (70%) of the survey respondents are from or reside in Switzerland.

Table 6 Professional work distribution (n = 245) of quantitative perception survey participants (N = 161)
Full size table
Table 7 Demographic distribution of quantitative perception survey participants (N=161)
Full size table

Table 8 shows the results of the quantitative perception survey for ten (out of 50) prominent examples of construction robots and common equipment with the full table in Appendix 3.

Table 8 Mean perception value per robot, range: 1–5 (full data in Appendix 3 and SD in the supplementary materials)
Full size table

Figure 4 consolidates the results of the perception of ease of use and work task adaptability. Note here that dimensions are qualitatively scaled to according min/max values and do not represent absolute values. The detailed table including standard deviation and variance can be found in the supplementary materials.

Fig. 4
figure 4

Visualisation of robots’ mean perceptions (x = Ease of use, y = Work task adaptability (N = 161), M. Baumann

Full size image

Figure 5 consolidates the results of the perception of ease of use and risk of job loss using the same visualisation approach.

Fig. 5
figure 5

Visualisation of robots’ mean perceptions (x =  Ease of use, y =  Risk of job loss (N = 161), M. Baumann

Full size image

Table 9 indicates low correlation between the attributes and thus proves the validity of using BLR as statistical tool for the analysis of our data set. The detailed descriptive statistic analysis can be found in Appendix 2.

Table 9 Correlation between robot attributes (n = 2213)
Full size table

Inferential statistics: All attribute expressions are then cross-compared in three different parameter calculation iterations ((1)–(3)): Morphology (1): Compares the indicator anthropomorphic to the coded variable of BLR iteration (1), which is zoomorphic. Morphology (2): Compares the indicator mechanomorphic to the coded variable of BLR iteration (2), which is anthropomorphic. Morphology (3): Compares the indicator mechanomorphic to the coded variable of BLR iteration (3), which is zoomorphic. Color (1): Compares the indicator bright to the coded variable of BLR iteration (1), which is dark. Material (1): Compares the indicator hard to the coded variable of BLR iteration (1), which is soft (Appendix 2).

The number of analyzed stimuli (n) is shown in the description of each table below. Note that the number of stimuli (n) is higher than the number of participants (N). This is because each participant judged multiple different pictures.

When computing the responses of all participants regarding the perception of ease of use, three significant (\(p<0.05\)) findings are supported by different stakeholder groups (n = 5): Exp(b) = 2.6 suggests that the odds of zoomorphic robots are 2.6 times higher to be judged as easy to use than robots in the anthropomorphic group. Exp(b) = 2.3 suggests that the odds of zoomorphic robots are 2.3 times higher to be judged as easy to use than robots in the mechanomorphic group. Exp(b) = 1.7 suggests that the odds of bright robots are 1.7 times more likely to be perceived as easy to use than dark robots (Appendix 2).

When computing the responses of all participants regarding the perception work task adaptability, two significant (\(p<0.05\)) findings are supported by different stakeholder groups (n = 5): Exp(b) = 1.7 suggests that the odds of anthropomorphic robots are 1.7 times higher to be judged as more multifunctional than robots in the mechanomorphic group. Exp(b) = 1.6 suggests that the odds of zoomorphic robots are 1.6 times higher to be judged as more multifunctional than robots in the mechanomorphic group (Appendix 2).

When computing the responses of all participants regarding the perception risk of job loss, three significant (\(p<0.05\)) findings are supported by different stakeholder groups (n = 5): Exp(b) = 2.0 suggests that the odds of anthropomorphic robots are twice as likely to be perceived as a risk to job loss than zoomorphic robots. Exp(b) = 1.9 suggests that the odds of mechanomorphic robots are 1.9 times as likely to be perceived as a risk to job loss than zoomorphic robots. Exp(b) = 1.5 suggests that the odds of dark robots are 1.5 times as likely to be perceived as a risk to job loss of human workers than robots that are bright (Appendix 2).

7 Key findings

Table 10 summarizes the study’s five key findings. They are not the only significant findings, but those are the findings that are validated by multiple stakeholder groups as well as the total sample group. The presented data of these studies show the results of the total participating sample group.

Table 10 Findings of the quantitative perception study
Full size table

While the results of most stakeholder groups point in the same direction, there are four exceptions among the different AEC stakeholders: Electrical and plumbing workers perceive bright robots 6.6 times more adaptable to work tasks than dark ones. Painting and glazing workers are seven times as likely to perceive anthropomorphic robots as adaptable to work tasks as mechanomorphic ones. Construction managers are 2.6 times more likely to perceive anthropomorphic robots as a risk to job loss than mechanomorphic ones. Painting and glazing workers are 7.2 times more likely to perceive anthropomorphic robots as a risk to job loss than mechanomorphic robots. These values are notably higher than previous stated key findings, yet there is a limited sample size in these stakeholder groups.

In addition, 191 participants from both surveys voluntarily answered an additional prompt on the difference between the terms robot and machine. The responses have been grouped qualitatively in the five following themes in Table 11. However, some respondents have stated that robots and machines are alike. More attention may be given to understanding these emerging themes.

Table 11 Themes of voluntary responses (N = 191)
Full size table

Further, 30 participants from the quantitative perception survey indicated prior experience working with robots in a second voluntary item. No a priori hypothesis on the participants’ former experience with robots was formed. We controlled for experience and could not find a significant difference of participants perception regarding their previous experience with robots. This is based on one-tailed, independent two sample homoscedastic t-tests (assuming equal variance) that result in: Ease of use: t(96) = 0.76, p = .23; Work task adaptability: t(96) = − 0.31; p = .38, Risk of job loss: t(96) = 0.37, p = .36. The detailed data analysis can be found in the supplementary materials. 

8 Discussion and limitations

Our study participant cohort may be positively skewed towards technological innovation due to their age: The participants were younger than the construction industry average in Germany (Gerlmaier and Latniak 2012). Further, the study had more female participation than the industry average in Germany (Hauptverband der Deutschen Bauindustrie 2022). We recognize these numbers are from a neighbouring country but expect similar averages across Germany, Austria and Switzerland. A validation study across more dispersed geographical locations should be conducted to find similarities and differences when studying construction robotics (e.g. Broehl et al. 2019).

Based on our study it is not clear what type of robotic system is considered a toy, device, tool, machine, robot or even a co-worker by the various stakeholder groups. In our study design we removed not only the context of the construction robot (such as the environment, control panels, human operator and brand names) but also its particular task allocation which might be critical in the scenario of a single task construction robot. The introduction of one or more specific construction tasks and collaboration scenarios will increase the understanding of user needs for such construction robots. Due to statistical considerations the study was limited to three attributes and three perceptions. We see the possibility to conduct a similar study design with more variables, given access to a bigger sample group. In addition, the use of continuous rating scales instead of Likert scales in online surveys has been proposed by scholars and could be implemented to further reduce statistical noise (Treiblmaier and Filzmoser 2011).

9 Conclusion and outlook

Our studies show that construction robot morphology has a statistically significant impact on these three perceptions: ease of use, work task adaptability and risk of job loss. Color has a statistically significant impact on the perception of risk of job loss: It is shown that brighter colors have a less negative impact on this perception. No analysis on the material can be expressed due to insufficient stimuli responses on this attribute. It is assumed that ease of use and work task adaptability are desired perceptions in robots, while the risk of job loss is an undesired perception. Our data also show that there is no superior morphology expression. In fact, some attributes of a single robot might perform well in one or two perceptions but fail the third. Depending on the particular requirements of the robot under study, the morphology should be determined individually. In addition, less represented stakeholder groups did not generate statistically significant relations between all attributes and perceptions.

Prospectively we envision the use of configurators as supplementary design tools to help conceptualize novel, modular robotic systems according to well-perceived visual attributes, beyond their purely technical engineering benchmarks. Scholars have coined the term Value Sensitive Design (VSD) to describe the incorporating values of ethical importance into engineering design (van de Poel 2013b). Therefore, the findings of this research study may support these user-centric, industrial design considerations for ongoing and future developments in the field of construction robotics. For example, into one or several versions of ETH’s “In-Situ Fabricator 2” concept (Giftthaler et al. 2017) or Rethink Robotics’ Baxter (2012) that is tailored to specific stakeholder groups in AEC—or perhaps one that is on average well perceived by all. However, to increase the likelihood of positive acceptance of robotics in the construction sector, the HRS must be deliberately designed. Additional socio-technical studies on the user’s expectations and acceptance, beyond visual perception, are required: For example, to extend or combine the research results with other, existing frameworks such the Unified Theory of Acceptance and Use of Technology (UTAUT, Venkatesh et al. 2003). Further human-centred work design principles such as job resources (levels of control and autonomy, skill use and task variety, job feedback, relational aspects) and job demands (e.g. performance monitoring) require direct integration in the development of the construction robots (Parker and Grote 2020).