Digital Technology Integration in Risk Management of Human–Robot Collaboration Within Intelligent Construction—A Systematic Review and Future Research Directions

Abstract

With the digital transformation of the construction industry toward intelligent construction, advanced digital technologies—including Artificial Intelligence (AI), Digital Twins (DTs), and Internet of Things (IoT)—increasingly support Human–Robot Collaboration (HRC), offering productivity gains while introducing new safety risks. This study presents a systematic review of digital technology applications and risk management practices in HRC scenarios within intelligent construction environments. Following the PRISMA protocol, this study retrieved 7640 publications from the Web of Science database. After screening, 70 high-quality studies were selected for in-depth analysis. This review identifies four core digital technologies central to current HRC research: multi-modal acquisition technology, artificial intelligence learning technology (AI learning technology), Digital Twins (DTs), and Augmented Reality (AR). Based on the findings, this study constructed a systematic framework for digital technology in HRC, consisting of data acquisition and perception, data transmission and storage, intelligent analysis and decision support, human–machine interaction and collaboration, and intelligent equipment and automation. The study highlights core challenges across risk management stages, including difficulties in multi-modal fusion (risk identification), lack of quantitative systems (risk assessment), real-time performance issues (risk response), and weak feedback loops in risk monitoring and continuous improvement. Moreover, future research directions are proposed, including trust in HRC, privacy and ethics, and closed-loop optimization. This research provides theoretical insights and practical recommendations for advancing digital safety systems and supporting the safe digital transformation of the construction industry. These research findings hold significant important implications for advancing the digital transformation of the construction industry and enabling efficient risk management.

1. Introduction

The rapid development of intelligent construction technologies is transforming traditional operational paradigms in the construction industry. Since the introduction of Building Information Modeling (BIM) in the 1970s, the sector has experienced technological evolution from digitalization to automation and. more recently, to intelligentization [1]. In recent years, the integration of digital technologies such as Artificial Intelligence (AI), Internet of Things (IoT), and Digital Twins (DTs) has created new HRC modes in intelligent construction contexts [2]. These technologies not only reshape conventional construction workflows but also show great potential in areas such as risk identification, safety monitoring, and early warning systems. However, construction remains one of the most hazardous industries worldwide. Since 2011, the construction industry has consistently recorded the highest number of occupational fatalities in the United States. In 2023, construction-related deaths accounted for nearly one fifth of all workplace fatalities, with a total of 1075 fatal injuries reported in the sector [3,4]. While HRC in intelligent construction environments enhances productivity and efficiency, it also introduces new risk management challenges, including physical space-sharing risks, cybersecurity threats, and systemic vulnerabilities [5,6]. Accordingly, applying digital technologies to manage these emerging risks has become a critical area of academic and practical concern.

Various digital technologies show promise in HRC-related risk management. IoT enables real-time monitoring and data acquisition on construction sites through intelligent equipment and sensing systems [2]. DT supports risk assessment through virtual simulations and real-time predictive modeling [7]. AI applications leverage machine learning algorithms for safety warnings and accident prediction by identifying latent risk patterns in historical data. Augmented Reality (AR) and Virtual Reality (VR) have been widely adopted for risk visualization and safety training, contributing to enhanced hazard perception and situational awareness [8]. Other emerging technologies such as collaborative robots (cobots), brain–computer interfaces, and exoskeletons have also been applied to reduce manual labor risks and support high-risk operational tasks [9,10]. However, the deployment of these technologies introduces additional risks, including cybersecurity concerns, system interoperability limitations, and increasing dependency on technological infrastructures [6,11].

Despite these developments, current research remains fragmented. Many studies focus on isolated technological applications and lack a comprehensive perspective on integrated risk management. As digital technologies become increasingly embedded in construction workflows, there is a pressing need for a structured framework to support coordinated application and address the multi-faceted challenges of intelligent construction. Therefore, this review systematically synthesizes the current state of research on the use of digital technologies for risk management in HRC within intelligent construction environments. The specific research objectives are as follows:

(1)

To analyze the current applications of digital technologies: identifying key technology types, use patterns, and effectiveness in HRC risk management.

(2)

To construct a digital technology system framework: mapping out the synergistic relationships among different technologies in HRC risk management.

(3)

To identify core challenges and future directions: including technical, managerial, and ethical issues, and to propose pathways for future development.

By providing a comprehensive review of recent studies, this study contributes to the theoretical understanding and practical guidance needed to advance digital safety solutions in intelligent construction, supporting the industry’s digital transformation and sustainable development.

This article is organized as follows: Section 2 describes the research methods of this study in detail; Section 3 illustrates results of the literature reviews; Section 4 presents the discussion and future work; Section 5 presents the results and limitations of this study.

2. Methods

This study employs a mixed method review approach that combines quantitative and qualitative analyses to minimize subjective bias and provide in-depth understanding of the knowledge structure and development trends within the research domain. The methodology comprises two phases: bibliometric quantitative analysis and systematic review qualitative analysis. The quantitative phase utilizes VOSviewer software (version 1.6.20) to conduct keyword co-occurrence analysis and construct knowledge mapping, revealing intrinsic relationships among research themes. The qualitative phase employs systematic review methodology to analyze research content and developmental trajectories in depth. The research framework is illustrated in Figure 1.

Based on three core concepts, intelligent construction, human–robot collaboration, and risk management, this study identified corresponding primary search terms and their synonyms. Considering the diversity in terminology usage, the search scope encompassed related terms (e.g., “Industry 4.0,” “Construction 4.0,” “health,” “hazard,” “accident”). The study employed Boolean logical operators (AND, OR) to combine keywords and utilized wildcard matching strategies to ensure comprehensive retrieval. The optimized search terms and search strings are presented in

Table 1. Keywords and search terms.

Keywords	Search Terms
A: Intelligent Construction	“Intelligent construction” “Smart construction” “Digital construction” “Automated construction” “Construction 4.0” “Industry 4.0 in construction”
B: Human–Robot Collaboration	“Human-machine collaboration” “Human-robot collaboration” “HRC,” “collaborative robots” “robots,” “man-machine interaction” “Construction robots” “Building robots”
C: Risk Management	“Safety management” “Worker safety” “Occupational safety” “Risk management” “Safety risks” “Construction safety”
Search Strings	TS = (((construct* OR build*) OR (intelligent* OR smart OR digital OR automat*)) AND ((human OR (machine* OR robot*)) AND collaborate*) AND (safe* OR risk* OR health* OR hazard* OR accident*))

.

This study used the Web of Science (WoS) Core Collection database for literature retrieval. As an authoritative academic database that indexes high-quality journal publications, WoS is widely adopted in scientometric research, and its comprehensive coverage of science and engineering fields ensures global representativeness of research findings [12]. The initial screening stage involved evaluation based on titles and abstracts using clearly defined inclusion and exclusion criteria. We considered studies eligible if they were published in English between 2010 and 2014, appeared as journal articles, and focused on safety management, risk management, or related topics. In addition, studies were required to involve a technical background associated with intelligent construction technologies (e.g., BIM, IoT, AI, robotics, etc.) or HRC. Then studies that were clearly irrelevant to the research scope were excluded based on a preliminary review of titles, abstracts, and keywords.

Articles that passed the initial screening advanced to the refined screening and quality assessment stages. In the refined screening, relevance was determined by examining whether the studies specifically addressed HRC issues, rather than general robotics or machines, and whether they provided insights into security risk characterization, management techniques, or prevention and control measures aligned with the research questions. The quality assessment then focused on evaluating the clarity of research objectives, the appropriateness of the methodology, the relevance of the findings, and the adequacy of reported details, thereby ensuring the robustness and reliability of the included studies.

3. Results

3.1. Literature Search Result

Based on the search strings, we retrieved a total of 7640 publications from the WoS Core Collection database. Literature screening constituted a core component of systematic reviews. This study adhered to the PRISMA [13] guidelines and employed a two-stage screening process to ensure the quality and relevance of included studies. The initial screening yielded 1129 articles, with 70 studies ultimately included following the final screening stage. The PRISMA [13] flow diagram for literature processing appears in Figure 2. In this step, no automation tools were involved (Please refer to the Supplementary Material for the PRISMA checklist).

3.2. Bibliometric Analysis Result

3.2.1. Annual Publications and Journal Distribution

Figure 3 shows the publication trends in intelligent construction, HRC, and risk management from 2010 to 2025. The development of this field can be categorized into two distinct phases: a slow initiation period (2010–2020) and a rapid growth period (2021–2025). From 2010–2020, the number of relevant publications remained limited, with only 15 articles published cumulatively, indicating that this multi-disciplinary field had not yet garnered widespread attention. In 2017, the field marked a turning point with literature quantities beginning to exhibit an upward trend. After 2021, the field entered a rapid development phase, with publication numbers surging from 4 articles in 2020 to 21 articles in 2024, representing an average annual growth rate of 67.8%. This significant growth primarily stems from the synergistic effects of the following factors: first, the optimization of the policy environment provided directional guidance for research development. Second, technological advances established the foundation for research practice. The large-scale deployment of 5G networks, mature applications of DT platforms, deep integration of BIM technology with robotic systems, and rapid development of construction robot hardware provided technical support for intelligent construction and HRC research. Third, the COVID-19 pandemic exposed the limitations of traditional construction models, with labor shortages and on-site safety challenges driving urgent demand for automation technologies, further stimulating growth in related research.

Overall, the publication volume in this field demonstrates distinct phased characteristics, transitioning from slow initial accumulation to rapid recent growth, reflecting the multiple effects of policy guidance, technological drivers, and market demand. The sustained growth in research interest indicates that intelligent construction, HRC, and risk management, as emerging interdisciplinary fields, possess significant theoretical value and practical importance.

An analysis of journal distribution in the fields of intelligent construction, HRC, and risk management reveals distinct interdisciplinary characteristics. Figure 4 presents the distribution of the top ten journals ranked by publication volume. IEEE Access leads with nine publications, demonstrating its academic influence in this interdisciplinary field. Advanced Engineering Informatics and Automation in Construction published six and five relevant articles, respectively, indicating sustained attention from the engineering informatics and construction automation domains to intelligent construction research. The journal distribution reflects the multi-disciplinary nature of the research, encompassing engineering, computer science, construction management, and safety science. IEEE Access, as an interdisciplinary journal, provides an important platform for intelligent systems and sensor technology research. Advanced Engineering Informatics focuses on engineering information technology applications, particularly BIM and data analysis, offering an ideal vehicle for intelligent construction research. Automation in Construction integrates mechanical engineering, computer science, and safety management, providing a comprehensive perspective for HRC safety research.

Based on 2023 academic data, high-impact journals exhibit distinct characteristics. Advanced Engineering Informatics (impact factor 8.0) and Automation in Construction (impact factor 9.6) demonstrate not only substantial publication volumes but also leading academic influence. Additionally, journals such as IEEE Transactions on Intelligent Transportation Systems (impact factor 7.9) and Expert Systems with Applications (impact factor 9.29) occupy important positions in the field. Notably, journals such as Sensors and Applied Sciences, despite having relatively lower impact factors (both 2.5), still provide important academic resources for researchers due to their open-access characteristics and broad disciplinary coverage, offering unique value in safety monitoring, environmental monitoring, and fundamental applied research.

Overall, the academic ecosystem of this field exhibits multi-disciplinary and interdisciplinary characteristics. IEEE Access, Advanced Engineering Informatics, and Automation in Construction constitute the core journal cluster, providing the most influential academic platforms for intelligent construction and HRC risk management research. This journal distribution pattern not only reflects academic attention to the field but also indicates its tremendous potential for future development.

3.2.2. Keyword Co-Occurrence

To comprehensively understand the main research streams and thematic distribution in this field, this study employed VOSviewer software (version 1.6.20) to construct a keyword co-occurrence network. Figure 5 shows the knowledge structure of intelligent construction, HRC, and risk management from 2010 to 2024, where node size reflects keyword frequency, internode distance indicates correlation degree, and different colors represent distinct research clusters.

Table 2. Keyword frequencies and total link strengths.

No.	Keyword	Frequency	Total Link Strengths
1	computer science	69	66
2	telecommunications	27	26
3	collaboration	17	16
4	digital twin	16	16
5	deep learning	16	14
6	framework	14	14
7	human-robot collaboration	14	14
8	management	14	14
9	system	14	14
10	construction	13	13
11	machine learning	12	12
12	model	13	12
13	design	11	11
14	federated learning	11	11
15	physics	12	11
16	materials science	11	10
17	safety	10	10
18	security	10	10
19	artificial intelligence	9	9
20	instruments & instrumentation	13	9

presents the occurrence frequency and total link strength of keywords. The proximity of “frequency” and “total link strength” values indicates that the associations among keywords are relatively direct, with few intermediary keywords.

The keyword co-occurrence analysis identified four major research clusters. The red cluster focuses on industrial technology and safety systems, centered on keywords such as “physics,” “Industry 4.0,” “automation,” and “safety,” embodying the deep integration of technological innovation and safety management in intelligent construction environments. High-frequency terms within this cluster, including “safety, operations research & management science,” reflect sustained attention to risk management and system performance optimization. The green cluster is dominated by intelligent technologies, with keywords such as “computer science, deep learning, machine learning and artificial intelligence,” highlighting the central position of AI technologies in safety management. The high frequency of “digital twin” and “framework” indicates a research focus shift toward constructing intelligent algorithm-based technical frameworks to enhance the predictive and intelligent capabilities of safety management. The blue cluster emphasizes HRC and digital technology integration, with keywords such as “human-robot collaboration, robotics, and augmented reality,” highlighting the role of emerging digital technologies in reshaping safety management paradigms. This cluster emphasizes that technology serves not merely as a tool but as a transformative force in reforming work environments. The yellow cluster focuses on technological application scenarios, with keywords such as “transportation, construction, and manufacturing” reflecting the expansion of research from specific industries to diverse application contexts. Notably, these clusters are interconnected through cross-domain keywords such as “system, performance, and design,” forming a complex knowledge network. The emergence of new technological terms such as “telecommunications” and “blockchain” further enriches the technological dimensions of research, transcending traditional disciplinary boundaries.

The keyword analysis reveals an evolutionary trend in this field from technology-driven approaches toward human-centered safety management methods, with research perspectives shifting from singular technological applications to complex socio-technical systems. This analysis provides a theoretical foundation for understanding the multi-disciplinary and interdisciplinary characteristics of HRC risk management in intelligent construction environments. Future research should further explore the synergistic effects among these clusters and their systematic improvements to workplace safety.

Based on keyword clustering analysis using VOSviewer (version 1.6.20), this study initially categorized literature keywords into four clusters: industrial technology and safety systems, computer science and intelligent technology, human–computer interaction and digital technology, and technology applications. Given the excessive homogeneity and overlapping themes within the initial clusters, this study conducted further refinement analysis by decomposing and reorganizing the labels, ultimately establishing nine sub-clusters consolidated into four primary dimensions.

Through analysis of the nine clusters, this study consolidates them into four primary dimensions: Digital Technology and Integration, Risk Management and Processes, Application Domains, and Future Development Directions. Figure 6 illustrates the reorganized clustering. Based on this dimensional framework, the systematic review will proceed in the logical sequence of digital technology and integration, application domains, and risk management and processes. The review first analyzes core digital technologies in the literature and their domain distribution characteristics, subsequently examines the practical implementation of these technologies across different application domains, and finally elucidates specific application patterns of these technologies across various stages of risk management.

3.3. Systematic Literature Review Result

3.3.1. Digital Technology and Integration

Among the reviewed literature, 97% of articles employ technological approaches to achieve worker safety management in Human–Robot Collaboration (HRC) environments. These technologies play crucial roles across different risk management stages. Through systematic analysis, four primary technology types emerge: Multi-modal Acquisition (MMA) technology, artificial intelligence learning technology (AI learning technology), Digital Twins (DTs), and Augmented Reality (AR).

Technological approaches constitute core elements of worker safety management in HRC environments, effectively enabling timely risk identification, accurate risk assessment, and scientific risk response, providing safeguards for worker safety. Based on this foundation, this paper will first conduct in-depth analysis of the characteristics and mechanisms of the four core technologies, followed by systematic review and critical analysis of the collected literature according to the four-stage risk management process.

(1)

MMA Technology

The configuration of MMA technology varies across application scenarios: medical contexts emphasize physiological signal fusion, while construction scenarios prioritize spatial awareness. Data acquisition refers to the process of collecting raw data from sensor networks (e.g., temperature, humidity, position) and transmitting them via wireless protocols (e.g., ZigBee, Wi-Fi, Bluetooth) for processing. Advanced sensors now include cameras and wearable devices for Electroencephalography (EEG) and Electrocardiography (ECG). To address the limitations of single-sensor systems in complex environments, multi-modal acquisition integrates diverse data sources (vision, motion, audio), enhancing perception and decision making in IoT, intelligent monitoring, and Human–Machine Interaction (HMI).

MMA technology refers to integrating multiple sensors or data sources (e.g., vision, sound, motion signals) to collect multi-dimensional information for fusion-based processing [14]. This approach enables combining heterogeneous data types, thereby overcoming the limitations of single-source data [15]. It has been widely supported in fields such as IoT, intelligent monitoring, and Human–Machine Interaction (HMI). Sensor fusion involving radar and cameras primarily supports environmental perception. For example, Ambrosino et al. [16] utilized a fused LiDAR–camera sensor system for environmental awareness and brick localization. Similarly, Chea et al. [17] integrated LiDAR and camera sensors to achieve environmental sensing and component localization, thereby ensuring assembly accuracy and minimizing geometric error risks. Kuru [18] employed LiDAR, camera, and radar systems to capture real-time road condition data. Du et al. [19] combined LiDAR, camera, and Inertial Measurement Unit (IMU) data and used the Robot Operating System (ROS) to process multi-source inputs—point clouds, images, and IMU signals—for generating high-precision 3D maps.

Wearable devices are mainly used for collecting human physiological and behavioral data, enabling predictive analysis and real-time response. Carsten and Martens [20] integrated sensor data (LiDAR and camera) with multi-modal information (visual, auditory, tactile) to ensure the effective transmission of critical information and reduce distraction risks. Chauhan et al. [21] applied wearable devices—Heart Rate Variability (HRV) monitors and Galvanic Skin Response (GSR) sensors—to collect workers’ psychophysiological data in real time and adjust robotic behaviors to enhance HRC safety and efficiency. Aasen and Klakegg [22] integrated camera data with EEG and ECG signals to improve driver monitoring system performance. Li et al. [23] used cameras, force sensors, and IMUs to collect real-time data and process multi-source information (e.g., location, force, posture) for motion prediction and path optimization. Liang et al. [24] combined RGB, depth, and skeletal data using depth sensors to extract spatiotemporal features and applied Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) to process multi-source data for human action recognition. Orsag and Koren [25] collected worker motion data using wearable IMUs, RGB-D cameras, and ultrasonic sensors to capture postures and 3D environmental information, enabling HRC distance detection and assisting in obstacle avoidance.

(2)

AI Learning Technology

AI learning technologies primarily encompass Machine Learning (ML), Deep Learning (DL), Reinforcement Learning (RL), and Federated Learning (FL). These technologies enable processing of complex data, protect privacy, and fulfill diverse task requirements in HRC. ML incorporates classical algorithms including random forest, Support Vector Machine (SVM), and decision tree. Becerik-Gerber et al. [26] employed ML algorithms to analyze building system data, effectively enhancing the response efficiency of intelligent lighting and HVAC systems. Montanaro et al. [27] utilized random forest algorithms to analyze RSSI signals for worker localization (achieving 1.5 m accuracy) and combined SVM and decision tree classification models for hazardous area intrusion detection. Chauhan et al. [21] applied random forest algorithms for data analysis to predict trust states, demonstrating ML’s advantages in complex data pattern recognition. Karmakar and Delhi [28] processed BIM data through ML techniques to achieve design optimization and risk prediction, showcasing broad prospects for cross-domain applications. DL primarily encompasses advanced methods such as Long Short-Term Memory (LSTM), Transformer, and Deep Reinforcement Learning (DRL). Cai et al. [29] leveraged LSTM and Transformer to process time-series data including posture and eye movement for human intention prediction. Liu et al. [30] employed the Rapidly exploring Random Tree Connect (RRT-Connect) algorithm to generate collision-free paths and optimized robotic arm path planning based on DRL algorithms to adapt to dynamic environments and complex component geometries. DRL includes methods such as A* algorithms and deep reinforcement learning. Chea et al. [17] and Gonzalez-Boehme et al. [31] applied A* algorithms combined with reinforcement learning to coordinate multi-robot actions, dynamically adjusting collaborative robot behaviors through real-time data processing. FL is typically combined with Differential Privacy (DP) and Secure Multi-Party Computation (SMPC). Arya et al. [32] directly addressed data protection requirements through federated learning and differential privacy technologies, effectively reducing cybersecurity risks in Vehicular Ad hoc Networks (VANETs). Cai et al. [29] employed differential privacy and secure multi-party computation to encrypt model updates, achieving secure collaboration with central aggregation servers.

Building upon traditional control algorithms and integrating high-dimensional data processing capabilities, various optimization and control algorithms targeting specific problems have emerged. Ambrosino et al. [16] integrated Model Predictive Control (MPC) algorithms into control architectures, optimizing robot motion through high-dimensional data processing, demonstrating deep integration between information technology and hardware systems. Du et al. [19] employed Simultaneous Localization and Mapping (SLAM) algorithms to generate preliminary maps in collaboration with data fusion modules. Pantano et al. [33] implemented dynamic task allocation based on genetic algorithms, achieving a 15% reduction in task execution time while balancing safety and efficiency. AI technology also proves highly effective in cost control applications. In the construction sector, Mansouri et al. [34] analyzed the application effects of BIM and machine learning in cost control, with research indicating that this technology combination can achieve 10% cost reduction.

(3)

Digital Twins

DT technology enables digital transformation in the construction industry. By establishing bidirectional real-time mapping between physical entities and virtual models, it integrates IoT, AI, and BIM technologies to achieve dynamic perception, simulation prediction, and decision optimization throughout the construction lifecycle.

Virtual–physical collaborative optimization enables deep integration between design and construction phases. Chea et al. [17] developed DT models based on BIM and integrated them into construction planning and monitoring systems to achieve real-time simulation of construction assembly processes. Eswaran et al. [35] employed DT technology to synchronize layout schemes with real-time data, automatically adjusting construction site layout plans based on real-time site conditions such as equipment position changes and material storage status. Dynamic construction scenario applications demonstrate advanced capabilities in complex environments. Li et al. [23] utilized DT technology to map physical environments through real-time data, constructed human and robot models in virtual spaces, and integrated cloud computing or edge computing platforms to form Human–Robot Collaboration Digital Twin (HRC-DT) systems. González-Böhme and Valenzuela-Astudillo [31] leveraged virtual models in DT technology to simulate robot behaviors, assisting craftsmen in predicting potential issues during collaborative processes. Real-time risk prediction and process optimization applications showcase significant safety and efficiency improvements. Liu et al. [36] constructed virtual models based on Unity3D (version: 2021.3.10f1c1) to real-time map human gestures, robot states, and environmental parameters in physical spaces, predicting gesture-guided collaborative outcomes through virtual spaces and optimizing robot actions to reduce collision risks. Marinelli et al. [37] combined DT with lean analytics, simulating construction scenarios based on BIM and real-time sensor data to predict waste and optimize HRC tasks, achieving 35% reduction in time waste through Value Stream Mapping (VSM) and DT identification and optimization of non-value-added activities. Liu et al. [38] developed virtual models of production lines using DT technology based on real-time sensor data to achieve fault simulation and process optimization. Pantano et al. [33] simulated factory environments and robot behaviors based on Unity or ROS-Gazebo platforms, updating DT data through IoT protocols such as MQTT to achieve safety risk management.

(4)

Augmented Reality

AR technology provides real-time guidance through AR glasses or mobile devices to assist complex task execution [39]. This technology demonstrates significant application value in HRC within the construction industry. Immersive environment construction typically involves AR technology working synergistically with VR technology. Bernasconi and Blume [40] employed VR and AR technologies to create immersive virtual building environments, supporting user experience testing and design validation. Eswaran et al. [35] utilized VR/AR technologies based on Unity/Unreal Engine platforms or HoloLens devices to create immersive environments while integrating BIM or production data. González-Böhme and Valenzuela-Astudillo [31] employed Mixed Reality (MR) technology, generating immersive visualizations through AR head-mounted displays (such as HoloLens) and VR environments, integrating with graphics processing units to achieve fusion of virtual information with real environments, providing craftsmen with intuitive operational interfaces and feedback mechanisms.

Training and visualization applications demonstrate significant effectiveness. Karmakar and Delhi [28] combined AR/VR with BIM, collaborating through display devices such as HoloLens to provide immersive visualization and training functions. Leite et al. [41] proposed applying VR/AR technologies to immersive safety training (such as virtual construction sites), enhancing immersion and personalized experiences. Mansouri et al. [34] evaluated the impact of VR in safety training, with research indicating a 20% reduction in accident rates.

On-site decision support enables real-time data visualization and alert functions. Liu et al. [38] visualized DT data and alert information through AR glasses (such as HoloLens), providing decision support for operators.

3.3.2. Application Domains

The 70 reviewed studies span four primary application domains: traditional industries dominate the literature, with 29 studies on construction and 25 on manufacturing; intelligent application domains include 8 studies on smart cities and 9 on autonomous driving.

In construction, the concept of Construction 4.0 highlights full-lifecycle HRC applications, from design to maintenance, applicable to diverse scenarios including residential, commercial buildings, and infrastructure [28]. Prefabrication, as a key process distinguished from traditional construction workflows, has become a focal area for robotic technology applications [42]. Specific applications of HRC include bricklaying [16,43,44], material handling [21,45,46], automated construction or assembly [17,30], surveying [19,47], framework construction [31], on-site collision avoidance [25,37,48,49,50,51], hazard monitoring [27,52,53], and worker training [34,54].

In manufacturing literature, HRC is applied across eight core processes: assembly, maintenance, training, quality control, mobility, machine operation, product and process design, and production planning and control [39]. This field proposes the “Operators 4.0” concept, referring to workers augmented by intelligent technologies capable of performing tasks more efficiently in intelligent manufacturing environments. Existing literature primarily focuses on assembly tasks [23,29,35,39,55,56,57], with limited coverage of equipment asset management [58,59,60,61], training [41], maintenance [62], and production optimization [38,63]. In smart cities, robotic technologies are primarily applied to building energy efficiency improvement [26], human settlement enhancement [40,64], intelligent monitoring [65,66], and smart city operations and maintenance [64]. The autonomous driving domain mainly concentrates on human–vehicle collaboration [67,68] and intelligent transportation systems [69].

Digital technologies related to HRC are predominantly applied in manufacturing and construction, while also demonstrating potential in daily life scenarios. The usage frequency of four core digital technologies varies significantly across domains. As illustrated in Figure 7, digital technology applications exhibit consistent trends across domains: AI learning technologies demonstrate the highest usage frequency across all fields, followed by MMA technologies, while DT and AR technologies are used relatively less frequently. This distribution pattern indicates that all domains place high emphasis on the application of multi-modal acquisition and AI learning technologies.

3.3.3. Risk Management and Processes

Risk management is a critical component of worker safety management. According to the Occupational Health and Safety Management Systems standard published by ISO [70], risk management comprises four main phases: risk identification, risk assessment, risk response, and monitoring and continuous improvement. This study analyzes the application of key digital technologies across these phases based on this framework.

Figure 8 presents the statistical distribution of digital technology applications across risk management phases. In terms of technology distribution, AI learning technologies account for 39%, maintaining a dominant position; MMA technologies represent 27%; while DT and AR technologies account for 19% and 15%, respectively.

Analysis by application phase reveals significant differences among technologies. In the risk identification phase, AI learning technologies dominate applications, reflecting the advantages of machine learning in pattern recognition and anomaly detection. MMA technologies are equally active in this phase, highlighting the importance of multi-source data fusion in early risk detection. In the risk assessment phase, both AI learning and MMA technologies exhibit excellent performance.

The technology distribution in the risk response phase is most significant. AI learning technologies show the highest application ratio, indicating the critical role of artificial intelligence in risk mitigation and response strategy formulation. MMA technologies perform prominently, demonstrating the importance of real-time monitoring in risk response. DT and AR technologies also show strong performance, revealing the value of virtual simulation and visualization technologies in solution testing and execution guidance.

In the monitoring and continuous improvement phase, all technologies show relatively low application frequencies, potentially reflecting the specific technological requirements of this phase or limitations of existing technologies in long-term monitoring. The following sections analyze the specific applications of each digital technology across the four risk management phases.

(1)

Risk Identification

Accurate risk identification is the prerequisite for safety management. Due to the complexity of construction scenarios, safety risk identification is characterized by complexity and data diversity. Figure 9a illustrates the digital technology applications in the risk identification phase. The integration of multiple technologies enables comprehensive risk detection and early warning capabilities.

MMA technology is extensively deployed for environmental perception and human state monitoring at construction sites. Implementation strategies include sensor fusion approaches combining cameras with LiDAR to achieve robust environmental awareness and wearable device integration for real-time physiological and behavioral data acquisition. These deployments enable comprehensive risk detection across multiple dimensions, from spatial hazards to worker fatigue states. AI learning technology plays a critical role in pattern recognition and anomaly detection. Applications focus on processing multi-modal sensor data through deep learning algorithms to identify potential safety risks in real time. The technology demonstrates particular effectiveness in complex scenarios where traditional rule-based methods prove insufficient. DT technology enables proactive risk identification through simulation-based analysis. By creating virtual replicas of construction environments, potential hazards can be identified and analyzed before they manifest in physical spaces. Askarpour et al. [71] identified high-risk scenarios through simulation and proposed corresponding countermeasures. AR technology supports risk identification through enhanced visualization and predictive analytics, providing workers with real-time hazard warnings and safety guidance.

(2)

Risk Assessment

In the risk assessment phase, digital technologies enable quantitative analysis and systematic evaluation of identified risks. Figure 9b illustrates technology applications in risk assessment. The integration of quantitative models and real-time monitoring enables systematic risk evaluation and informed decision making.

AI learning technology serves as the core analytical engine. ML models, particularly random forest and DL algorithms, quantify risk probabilities and severity levels based on historical data and real-time inputs. Orsag and Koren [25] achieved 88% accuracy in collision risk quantification using random forest models, while Pantano et al. [33] demonstrated 90% accuracy in risk probability assessment. These applications showcase the transition from empirical judgment to data-driven decision making. MMA technology provides continuous data streams essential for dynamic risk assessment. Multi-sensor fusion enables real-time monitoring of both environmental conditions and human factors, supporting adaptive risk evaluation as conditions evolve. DT technology facilitates scenario-based risk assessment through virtual simulation. System state verification and what-if analysis can be performed in digital environments before implementing changes in physical spaces. Perera et al. [72] used BIM to simulate fire spread and evacuation paths, assessing risk severity and demonstrating the value of virtual simulation in emergency response optimization.

(3)

Risk Response

Risk response is the process of changing risk characteristics through selecting and implementing measures. According to NIST AI risk management frameworks [73], risk response strategies include risk avoidance, mitigation, transfer, and acceptance. From an execution logic perspective, they are divided into preventive response and real-time response measures. Figure 9c illustrates the digital technology applications in this phase.

AI learning technology is the core technology for risk response. Risk prediction uses algorithms to predict human intentions or actions, adjusting robot behavior in real time to avoid risks. Chauhan et al. [21] employed ML models to predict trust levels, establishing a “data collection, trust analysis, model prediction, feedback optimization” loop to reduce collaboration risks. Ding and Zhou [74] used ML algorithms to identify potential risks and generate early warning signals. Li et al. [23] employed DT technology to analyze human postures and used AI to predict human behavior for optimizing robot actions. Karmakar and Delhi [28] used BIM digital management to reduce design errors and predicted construction risks through AI. Orsag and Koren [23] used random forest to predict collision risks and employed A* algorithms to adjust robot paths in real time, reducing collision risks by 65%.

Real-time response dynamically adjusts routes through on-site real-time information to prevent collisions. Liu et al. [30] corrected robotic arm deviations through real-time point cloud updates and dynamically adjusted paths. Liu et al. [75] used RGB-D cameras and LiDAR to generate real-time 3D maps, coupled with force sensors to correct robot actions. Liu et al. [43] used real-time physiological signals from wearable devices and employed reinforcement learning to dynamically adjust robot behavior, reducing safety accident risks by 40%. Montanaro et al. [27] employed RFID positioning and used ML models for intrusion detection (96% accuracy), reducing accident risks by 45%.

Management control serves as preventive response measures, including safety training and personal protective equipment. Ko et al. [45] reduced wearing fatigue through lightweight exoskeleton design, decreasing muscle load by 30%. Leite et al. [41] applied VR and AR for immersive safety training to enhance worker safety awareness.

(4)

Risk Monitoring and Continuous Improvement

Risk monitoring and continuous improvement represents the terminal phase of the risk management process, constituting the dynamic control core of the Occupational Health and Safety Management System (OHSMS). This phase achieves spiral enhancement of risk control through data-driven closed-loop management. Figure 9d illustrates the digital technology application framework for this stage.

This phase primarily employs real-time monitoring technologies to continuously collect on-site data, utilizes blockchain distributed ledger for data certification and document management, and implements systematic improvements through DT, AR, and VR technologies. Karmakar and Delhi [28] integrated blockchain with supply chain systems to support smart contracts and transparent data sharing, combining cloud computing and edge computing technologies for seamless connectivity with on-site equipment and ERP systems to achieve comprehensive construction process management. Montanaro et al. [27] deployed ML models at construction sites to dynamically update positioning and detection algorithms, continuously tracking worker locations and safety status through IoT networks and dashboards, achieving hazard response times below 1 s. Pantano et al. [33] employed DT and sensor systems to continuously monitor site conditions, performing real-time risk assessment updates and systematic risk event recording. Pinto et al. [52] identified environmental risks through 3D point cloud and image detection technologies, continuously collecting high-density point cloud data and comparing historical records to monitor structural changes, achieving a 30% improvement in inspection coverage through dynamic path adjustment.

4. Discussion and Future Research

Currently, the development of digital technologies brings valuable convenience to risk management in intelligent construction HRC, but there are still many challenges that mainly present in the following aspects: difficulties in multi-source heterogeneous data fusion [76], insufficient generalization capability of algorithms [77], lack of real-time performance in on-site management [78], and the lack of unified standards [79,80]. These challenges are reflected across various risk management stages. Next, the challenges faced by current technological development will be elaborated from the different stages of risk management.

4.1. Risk Identification

As the starting point of risk management, the comprehensiveness of risk identification determines the accuracy and efficiency of subsequent stages. Failure to identify major risks could lead to assessment deviations or the ineffectiveness of response strategies. The risk identification stage primarily relies on sensor-collected data for detection. Due to the complexity of intelligent construction environments, single-modal data cannot systematically reflect on-site conditions, potentially resulting in risk oversight. Although current research has begun adopting multi-modal technology, it predominantly uses vision monitoring or wearable devices for risk identification, which still have significant limitations. For example, visual modalities fail in low-light or occluded scenarios and environmental interruption (light, temperature) reduces the accuracy of sensor data from eye tracking and Heart Rate Variability (HRV) [80].

Overall, the applications of multi-modal technology have shown advantages over single-sensor approaches in risk identification regarding information collection, exhibiting higher accuracy [81] and superior complex scenario parsing capabilities [82]. However, challenges remain in modal heterogeneity, insufficient connectivity, and inadequate interactivity. Munikoti et al. [83] mentioned that multi-modal technologies mainly focus on traditional modes including images, texts, and videos, and there is limited research on emerging modes. Additionally, Generative Multi-modal Models (GMMs) face issues of poor scalability and insufficient benchmarking. Moreover, application of multi-modal technology usually requires integration with other systems (e.g., BMSs), which leads to complexity. For example, digital twin solutions depend on robust data acquisition systems, and their integration with existing Building Management Systems (BMSs) requires specialized engineering support [84]. Challenges include integrating sensor data, log data, and human–robot interaction data, latency and accuracy issues, difficulties in fusing multi-source heterogeneous data, and low sharing and storage efficiency due to the lack of unified data models.

4.2. Risk Assessment

Risk assessment is the process of quantifying risks and measuring potential losses. This stage plays a crucial guiding role in the adoption of risk response measures. With the development of technology, risk assessment has gradually evolved from experience-based judgment to data-driven decision making (e.g., CNN vision identification). However, quantitative risk evaluation and graded response remain insufficient. Risk management and control in most literature primarily follow an “identification–prediction–response” approach, which skips the risk assessment stage [85]. Consequently, it is difficult to conduct appropriate risk assessment and implement suitable response measures.

Furthermore, there are more new risks distinct from traditional construction in the intelligent construction field. Current risk assessment relies more on AI learning technologies, such as deep learning and machine learning models. These technologies demonstrate good accuracy in identifying and assessing some generic risks but lack sensitivity towards specific risks (e.g., workspace availability, off-site personnel safety) and real risks [86]. In contrast, humans are more adept at identifying context-dependent risks based on implicit knowledge and project experience and are also more accurate when analyzing project-specific risks (e.g., unique terrain constraints). Therefore, there should be more attention concentrated on the accuracy while using algorithms for risk assessment.

4.3. Risk Response

Risk response is the most important stage within risk management processes, and its success directly determines the outcome of risk management. Risk response primarily includes risk prediction, real-time risk response, as well as personal protection and training of workers. Risk prediction generally uses AI algorithms and other methods to analyze identified information and make predictions to take measures. Technologies like digital twins and BIM simulation are also used for risk prediction. However, many of these predictions are only simulations and have not been validated in real-world settings [87]. Especially for digital twins, a difference between synthetic data and real data exists. Construction sites are dynamically changing, and this discrepancy may affect the model’s generalization capability, making it difficult to ensure the practicality of proposed measures [88].

For real-time risk response, due to the massive volume of information and the enormous computational load of algorithmic models, substantial computational costs arise. This leads to the problem of model calculation delays [89], which is a significant obstacle to real-time risk response. Furthermore, workers exhibit individual differences in the areas of worker personal protection and training, and there is the challenge of lacking personalized countermeasures [90]. For example, when learning to use AR devices or other technical skills, workers have varying proficiency levels. The sensitivity and adaptability of models for conducting analyses need improvement.

4.4. Risk Monitoring and Continuous Improvement

Risk monitoring and continuous improvement ensure the effectiveness of response measures through regular reviews and feedback mechanisms and dynamically adjust strategies according to environmental changes. Monitoring results will be fed back to the risk identification and assessment stages, forming closed-loop management. The techniques used in this stage are like those in the risk identification stage (such as multi-modal data collection and digital twins) but also include various analytical algorithms like AI learning for strategy adjustment. In current research, continuous monitoring can be achieved through vision sensors and digital twins, etc., but continuous improvement has not been notably demonstrated in existing research. Research often focuses on real-time risk prediction and response [91,92]; studies on strategy adjustment through feedback from risk response measures are relatively scarce [93]. Risk continuous improvement requires establishing tracking and feedback mechanisms to promptly resolve issues and adjust strategies, forming a closed-loop process. However, the application of intelligent construction technologies in risk continuous improvement still faces challenges such as data security, technological updates, and personnel training, requiring further exploration.

4.5. Towards an Integrated Digital Approach to HRC

The primary technology was reviewed in the previous section, and the interrelationships between each technology were also clarified. Multi-modal acquisition technology provides various data input for AI learning technology, then the analysis result from AI learning technology also provides modeling indicators and optimization strategies. The simulation results of DT technology are visually presented to users via AR, thus forming a closed-loop technological workflow. The cloud–edge collaboration technology permeates the entire framework, ensuring rational allocation of computational resources and real-time system responsiveness. The digital technologies framework of HRC in intelligent construction was developed based on four key digital technologies and contains five layers. Figure 10 illustrates this HRC digital technology framework.

The bottom layer is for data perception and acquisition, with the function of collecting, monitoring, and analyzing real-time data from the construction site. By applying multi-modal acquisition technology, comprehensive data collection from the environment to people becomes achievable. These diversified data sources deliver rich and precise informational inputs for higher-level intelligent analytics. The layer of data transmission and storage aims to support data communication and remote collaboration at construction sites. Through coordinated collaboration across infrastructure, enabling technologies, and service levels, this layer ensures low-latency characteristics of data transmission channels while guaranteeing global consistency and accuracy of data storage. The function of the intelligent analysis and decision support layer is to analyze construction data, provide decision support, and optimize strategies. Technologies such as machine learning, deep learning, reinforcement learning, and differential privacy can help identification and analysis of text and images, as well as support construction forecasting and process optimization. The integrated multi-algorithm architecture helps the system adapt to complex on-site analysis requirements, which plays an important role in responding to risks at construction sites. Moving to the upper layers, the layer of human–robot interaction and collaboration focuses on supporting the interaction among construction workers, engineers, and intelligent devices. Integrated through a unified building management system, this layer enables command dispatch, remote operations, real-time monitoring/simulation, and equipment manipulation, while ensuring operational transparency to enhance HRC trust. The top layer, intelligent equipment and automation, supports automated construction processes, equipment control, and robot operation, which includes construction robotics, Unmanned Aerial Vehicles (UAVs), and IoT-enabled machinery. As the physical manifestation of digital technologies, it executes construction tasks and performs real-time monitoring/alerting, with collected terminals’ data feeding back into the data acquisition and perception layer to enable continuous optimization.

This framework embodies a complete technology ecosystem from data-driven intelligence to automated execution, establishing a closed-loop workflow of “perception–transmission–analysis–interaction–execution.” Through multi-layer technology integration, it effectuates a paradigm shift from passive monitoring to proactive governance and from fragmented applications to systemic integration. Moreover, the feedback loop from execution to perception layers enables adaptive learning and continuous refinement. The digital technologies framework of HRC proposed in this study can provide systematic technical solutions for intelligent construction and can also offer valuable reference models for the digital transformation of other complex systems.

Beyond its practical applications, this framework advances HRC theory in three significant ways. First, it transcends existing component-level reviews by proposing a holistic theoretical model that systematically integrates multi-layered technologies into a unified ecosystem, addressing the gap in comprehensive frameworks by bridging perception, cognition, and action in construction HRC. Second, the framework theorizes a paradigm shift from passive monitoring to proactive governance, establishing a closed-loop workflow that moves beyond fragmented, reactive applications toward systemic, anticipatory integration. This conceptualizes HRC as an orchestrated multi-layer process rather than isolated technological interventions. Third, the layered architectural approach provides a generalizable theoretical template for digital transformation across complex systems beyond construction, contributing to broader cyber–physical systems and Industry 4.0 knowledge.

Along with the proposed framework, future research should address emerging challenges across three interconnected dimensions that correspond to the “human–robot–system” paradigm of the framework. Key areas to explore include.

(1)

Human-Centered Research: Trust and Acceptance: Enhance transparency through user-friendly interfaces and quantify human trust levels to improve worker acceptance and system reliability in HRC.

(2)

Robot-Level Research: Multi-Modal Data Fusion: Solve multi-source heterogeneous data fusion challenges by establishing unified standards and optimizing algorithms for improved efficiency and accuracy of intelligent analysis.

(3)

System-Level Research: Privacy, Ethics, and Risk Management: Address privacy concerns through encryption and anonymization while ensuring AI fairness and regulatory compliance in construction data management. Transform risk management from experience-dependent to data-driven through automated learning, real-time monitoring, and intelligent feedback mechanisms.

5. Conclusions

This hybrid literature review systematically integrates interdisciplinary research across intelligent construction, HRC, and risk management domains, examining digital technology applications and challenges in HRC scenarios to inform industry safety management. The analysis reveals that core digital technologies, multi-modal acquisition technology, AI learning technology, DTs, and AR, demonstrate concentrated deployment patterns across risk management stages, with AI learning and multi-modal acquisition technologies assuming pivotal roles. Risk management approaches are transitioning from passive identification toward proactive response strategies, with traditional high-risk industries favoring mature technologies while emerging sectors prioritize innovative applications. Through keyword co-occurrence analysis and literature synthesis, this study establishes a digital technology framework for intelligent construction HRC, which provides systematic guidance for technology implementation and offers a replicable model for complex system digitalization. However, these findings are derived from a limited sample of reviewed studies and should be interpreted with appropriate caution.

Current implementations face significant limitations: single-modal data constraints, multi-modal fusion complexities, system integration challenges, absent quantitative frameworks, insufficient predictive model validation, and inadequate feedback mechanisms—collectively restricting performance in dynamic environments. Future research priorities include advancing multi-modal fusion and integration standardization for enhanced data precision and timeliness; developing quantitative assessment frameworks and real-time response models for emerging risks while improving AI generalization capabilities; strengthening trust mechanisms, privacy safeguards, and ethical protocols in HRC through personalization and lightweight solutions; and establishing closed-loop risk management systems with robust feedback and continuous improvement mechanisms. While this interdisciplinary synthesis provides a foundation for theoretical advancement and practical innovation in intelligent construction, the findings should be considered preliminary given the study’s scope limitations and the evolving nature of digital technologies in this domain. Further empirical validation through larger-scale studies and real-world implementations is necessary to strengthen these conclusions and address the complex implementation challenges in digital technology deployment.