Autonomous Earthwork Machinery for Urban Construction: A Review of Integrated Control, Fleet Coordination, and Safety Assurance

Abstract

Autonomous earthwork machinery is gaining traction as a means to boost productivity and safety on space-constrained urban sites, yet the fast-growing literature has not been fully integrated. To clarify current knowledge, we systematically searched Scopus and screened 597 records, retaining 157 peer-reviewed papers (2015–March 2025) that address autonomy, integrated control, or risk mitigation for excavators, bulldozers, and loaders. Descriptive statistics, VOSviewer mapping, and qualitative synthesis show the output rising rapidly and peaking at 30 papers in 2024, led by China, Korea, and the USA. Four tightly linked themes dominate: perception-driven machine autonomy, IoT-enabled integrated control systems, multi-sensor safety strategies, and the first demonstrations of fleet-level collaboration (e.g., coordinated excavator clusters and unmanned aerial vehicle and unmanned ground vehicle (UAV–UGV) site preparation). Advances include centimeter-scale path tracking, real-time vision-light detection and ranging (LiDAR) fusion and geofenced safety envelopes, but formal validation protocols and robust inter-machine communication remain open challenges. The review distils five research priorities, including adaptive perception and artificial intelligence (AI), digital-twin integration with building information modeling (BIM), cooperative multi-robot planning, rigorous safety assurance, and human–automation partnership that must be addressed to transform isolated prototypes into connected, self-optimizing fleets capable of delivering safer, faster, and more sustainable urban construction.

1. Introduction

In recent years, the construction industry has undergone a significant transformation with the integration of autonomous earthwork machinery, particularly in urban settings where space constraints and safety concerns are paramount [1]. This shift is driven by the need to address the growing demands of urbanization, where construction projects frequently occur in densely populated areas with limited space and proximity to existing structures [2,3]. The adoption of autonomous machinery such as excavators, bulldozers, and loaders is increasingly recognized as a solution to enhance operational efficiency and ensure safety in these complex environments [4]. According to research and markets’ Autonomous Construction Equipment Market Report 2025, the global autonomous construction equipment market is expected to reach USD 15.13 billion by 2025, with a compound annual growth rate (CAGR) up to 11.1%, reflecting the industry’s push toward improved productivity and safety [5]. This growth underscores the limitations of traditional construction methods in urban contexts, where advanced technologies are becoming indispensable.

Urban construction presents distinct challenges that necessitate innovative technological interventions. Limited space restricts the maneuverability of heavy machinery, while proximity to existing buildings and infrastructure heightens the risk of damage or accidents [6]. Additionally, high pedestrian traffic in urban areas amplifies safety concerns, making it critical to protect both workers and the public [7]. These combined pressures call for solutions that shorten schedules, curb material waste, and maintain rigorous safety margins. Integrated control systems that fuse high-accuracy positioning, multi-modal sensing, and AI-driven decision-making have emerged as a key enabler of such goals [8,9,10]. By coordinating heterogeneous fleets in real time, these systems sustain efficient, collision-free operation even within tightly confined urban work-zones.

To clarify the terminology used in this paper, the following definitions are provided, supported by academic literature:

Autonomous emphasizes full autonomous decision-making ability, while automated focuses on preset program control; recent survey work provides a comprehensive taxonomy of such systems [1].

Integrated control systems are cyber–physical architectures that combine sensor data, AI reasoning, and wireless communication to orchestrate multiple machines or process stages; recent prototypes range from shared-control overlays to IoT-enabled data hubs [10].

Risk-mitigation strategies comprise methods for identifying, analyzing, and controlling hazards posed by autonomous equipment, including predictive safety monitoring of robotic excavators [7].

Fleet interoperability denotes the ability of multiple autonomous machines to work together in a coordinated manner, sharing information and adapting to each other’s actions to achieve common construction goals efficiently and safely, facilitated by integrated control systems which explore internet of things (IoT) and artificial intelligence (AI) integration in construction automation [11].

Accordingly, this review undertakes a systematic synthesis of 157 peer-reviewed studies to answer two overarching questions: (i) how recent advances in perception-driven autonomy, integrated control architectures, safety and risk-mitigation strategies, and fleet-level interoperability are reshaping earthmoving practice on space-constrained urban sites; and (ii) what technical, managerial, and regulatory gaps must still be bridged before connected, self-optimizing equipment fleets become routine. After describing the search protocol and analytic methods in Section 2, we present descriptive statistics and bibliometric maps that locate the intellectual growth and collaborative structure of the field in Section 3. Section 4 then integrates these quantitative patterns with a qualitative content analysis to discuss the four thematic pillars in depth: autonomous machinery, integrated control, safety assurance, and multi-machine coordination. Building on the identified research gaps, Section 5 distils five cross-cutting priorities that define the agenda for the next decade, ranging from adaptive AI and digital-twin integration to rigorous safety certification and human–automation partnership models. Finally, Section 6 summarizes the main insights and outlines their practical implications for researchers, equipment manufacturers, and urban contractors.

2. Materials and Methods

A comprehensive literature search was conducted in the Scopus database to identify research on autonomous and automated earthwork machinery. The search was performed on 25 March 2025, covering the 10-year range from 2015 through early 2025. The query was formulated using combinations of keywords related to autonomy (e.g., autonomous and automated), construction machinery (e.g., earthwork, excavation, and equipment), and integrated control or safety (e.g., control systems and risk mitigation). For example, one search string used was (autonomous OR automated) AND (earthwork OR construction) AND machinery, alongside parallel queries targeting integrated control systems and safety (e.g., (integrated OR centralized) AND control AND systems AND construction, and risk AND mitigation AND (construction OR earthwork)), among others.

Table 1. Search String on Scopus.

Search String

TITLE-ABS-KEY ((((autonomous OR automated OR robotic) AND (excavator * OR dozer * OR loader * OR earthwork OR “heavy machinery” OR “construction equipment”)) OR (“integrated control system” OR “machine control” OR “fleet management system” OR “fleet interoperability” OR “multi-machine coordination”)) AND (construction OR “civil engineering”)) OR TITLE-ABS-KEY(((“risk assessment” OR “risk mitigation” OR “safety standard” OR “safety protocol”) AND (autonomous OR automated OR robotic OR “integrated control system” OR “machine control” OR “fleet management” OR “fleet interoperability”)) AND (construction OR earthwork)) AND PUBYEAR > 2014 AND PUBYEAR < 2026 AND (LIMIT-TO (SRCTYPE, “j”)) AND (LIMIT-TO (DOCTYPE, “ar”) OR LIMIT-TO (DOCTYPE, “re”)) AND (LIMIT-TO (LANGUAGE, “English”)) AND (LIMIT-TO (SUBJAREA, “ENGI”))

Note: The asterisk “*” indicates a wildcard to match all search terms that start with a certain word.

illustrates our search string combination on Scopus.

The initial Scopus query returned 597 documents. Titles and abstracts of these records were then screened to remove works outside the scope of construction automation and earthmoving machinery. In this screening, studies clearly unrelated to the construction domain were excluded. For instance, papers on biomedical engineering or generic project management, as well as those focusing on agricultural or mining machinery were set aside. We also excluded studies dealing with purely manual equipment operation without any automation component. After this filtering relevance, a final set of 157 publications was identified as directly pertinent to autonomous earthwork machinery and integrated control in construction. For each of these remaining papers, bibliographic data (title, authors, year, journal, keywords, etc.) were compiled for analysis.

A bibliometric analysis was then carried out on the curated dataset of 157 papers. We employed VOSviewer (version 1.6.20) to construct and visualize bibliometric networks. VOSviewer is a software tool that enables the mapping of co-authorship networks, co-occurring keywords, and other relationships from scholarly data. Using this tool, we generated (1) a keyword co-occurrence network to reveal the major research themes and their interconnections, and (2) an author collaboration network to identify influential authors and collaborative groupings, and extracted bibliometric indicators, such as (3) the yearly publication trend in the field, (4) the distribution of publications by journal, and (5) the geographical distribution of research output by authors’ countries [12]. The co-occurrence map was based on author-provided and index keywords in the Scopus records, where nodes represent frequently used keywords and links indicate their co-occurrence in the same publications. The co-authorship map was similarly constructed, with authors as nodes linked by edges if they co-authored at least one paper together. These networks were processed such that only significant nodes (e.g., keywords appearing in multiple papers, authors with multiple publications) were included to improve interpretability. The resulting visual maps were examined for clusters—groups of nodes more strongly connected to each other than to the rest of the network—which indicate subtopics or collaborator communities within the field. The bibliometric figures were prepared with appropriate annotations to support the narrative analysis. In the next section, we present the results of this bibliographic analysis, highlighting publication trends, prominent research themes, leading journals, and the collaboration patterns underlying the domain of autonomous earthwork machinery.

3. Bibliometric Analysis

As shown in Figure 1, the field has continued its rapid expansion. In 2015 only a single paper dealing explicitly with autonomous and robotic earth-moving was indexed. Annual output stayed in single digits through 2018 but climbed sharply thereafter: 17 papers in 2019 and 22 in 2020. The pace then stabilized at a little over 20 publications a year (21 each in 2021 and 2022, and 18 in 2023) before another surge to a decade-high 30 papers in 2024. Even the partial count of three papers for January–March 2025 suggests that the current year will again finish well into the double digits once the full record is available. Overall, annual production has risen thirty-fold over the past decade—clear evidence of escalating academic and industrial interest as machine guidance, on-board sensing, and AI-based control find their way into construction fleets.

The research outputs remain distributed across both construction–engineering and core robotics venues, underscoring the field’s interdisciplinary character. As illustrated in Figure 2, Automation in Construction is by far the dominant outlet with 38 papers (≈19% of the corpus), confirming its role at the construction-automation nexus. It is followed by IEEE Robotics and Automation Letters (10 papers), the broad interdisciplinary IEEE Access (8), and two further titles that straddle civil engineering and mechatronics—the Journal of Robotics and Mechatronics (6) and the Journal of Computing in Civil Engineering (6). Specialist robotics venues, such as the Journal of Field Robotics (5), and general applied-science journals, such as Applied Sciences (Switzerland) (4) and IEEE/ASME Transactions on Mechatronics (4), round out the top eight. This spread shows that findings are disseminated simultaneously to construction practitioners and the robotics community, facilitating cross-fertilization between civil engineering problems and control/mechatronic solutions.

Country counts are based on at least one author affiliation from the country concerned (Figure 3). China leads decisively with 45 papers, reflecting its massive infrastructure pipeline and government push for “smart construction.” Korea (31) has overtaken the United States to claim second place—an outcome that tracks long-running national programs in construction robotics spearheaded by KAIST and other institutes. The United States (27) remains a major contributor through university labs that focus on productivity, safety, and heavy-equipment automation. Canada (17) and Japan (14) form the next tier, each hosting active research clusters (e.g., trajectory optimization at University of Alberta; field robotics at Osaka and Tokyo universities). Europe’s engagement is led by Germany (8) and Switzerland (7), with Australia and Sweden (5 each) and Finland (4) completing the top ten. The pattern points to a globalized yet regionally clustered research landscape: East Asia anchors the volume, North America supplies a comparable breadth of topics, and selected European groups provide depth in sensing, control, and mechatronic design.

Figure 4 depicts the keyword overlay network for the 157 articles, with node color representing the average year of publication (dark-blue ≈ 2019 shifting to light-yellow ≈ 2023). At the center of the map lie construction equipment, excavation, and automation. These high-frequency terms form the structural backbone of the field and appear in green tones, indicating that they have remained continuously active from the outset of the review period through to the most recent papers. The density of links radiating from this core confirms that almost all later developments continue to anchor their contributions in problems associated with automated earth-moving machinery.

Moving outward from the core, the color gradient traces a clear temporal progression in research emphasis. Nodes situated in the blue region—such as hydraulic actuators, sliding-mode control, and trajectory control—belong predominantly to work published around 2019. These studies concentrate on low-level motion and force regulation of plants, reflecting a phase in which researchers sought to stabilize and precisely command the electro-hydraulic subsystems of excavators, loaders and similar equipment. Surrounding this early control layer is a band of green nodes, including computer vision, object detection, and convolutional neural networks. Dated to 2020–2021, these keywords mark the rapid adoption of deep-learning-based perception modules that deliver real-time scene understanding and target recognition; their strong edges to the core machinery terms demonstrate how sensing was swiftly integrated with traditional control architectures.

The outermost zone is dominated by yellow nodes, such as reinforcement learning, energy efficiency, and optimization, together with application-specific terms, like wheel loaders and piles. Averaging around 2021–2022, these keywords reveal the most recent shift in attention: away from merely controlling individual machines towards higher-level decision-making, multi-machine coordination, and sustainability objectives. In particular, the juxtaposition of reinforcement learning with productivity and trajectory planning suggests that authors are now tackling site-scale scheduling and path optimization problems using data-driven policies; meanwhile, the emergence of energy efficiency indicates that carbon and fuel considerations are being embedded directly into autonomy algorithms rather than treated as external constraints.

Taken together, the overlay visualization portrays a field that has matured from equipment-level control foundations to perception-augmented operation and is now advancing into optimization and environmental performance. The chronological layering highlights a natural autonomy stack—hardware control, environmental perception, and strategic planning—yet also exposes the research frontiers that demand further exploration. Robust, transferable reinforcement-learning schemes for heterogeneous fleets, frameworks that co-optimize task execution with energy or carbon metrics, and large-scale, real-world data sets for benchmarking remain comparatively sparse. Addressing these gaps will be essential for turning the demonstrated technical feasibility of autonomous earth-moving into dependable, sustainable practice on construction sites.

Figure 5 depicts the co-authorship network that emerges when a productivity threshold of three documents per author is applied to the 157-paper corpus. Twenty-nine authors satisfy this criterion, yielding a graph with nine distinct collaboration clusters (Items = 29; Clusters = 9; Links = 41; total link strength = 124). In VOSviewer, link strength equals the number of papers co-authored by a given pair; the sum over all edges therefore quantifies the overall intensity of collaboration. Node size is proportional to each author’s publication count, whereas link thickness expresses the intensity of shared output. The layout is computed with a force-directed algorithm in which the attraction factor controls how strongly linked nodes are pulled together, while the repulsion coefficient sets a global push that prevents node overlap [13]. To improve readability the layout was recalculated with an attraction factor of 2.0 and a repulsion coefficient of −1, thereby compressing intra-cluster spacing while retaining the original modularity. With these parameters the network achieves a modularity of Q = 0.79, indicating a clearly articulated but not over-segmented community structure. Modularity values above roughly 0.30 are generally interpreted as evidence of well-defined clusters [14].

The largest cluster in green is led by Feng Hao, Yin Chenbo, and Cao Donghui, whose joint output on redundant-sensor fusion and safety envelopes for autonomous earthmoving fleets accounts for almost one-third of the entire dataset. Their early work established a controls foundation: a genetic algorithm (GA)-tuned PID scheme for precise trajectory tracking [15] was followed by the systematic identification of Stribeck-curve friction in 2019 [16,17]. Building on these models, the team proposed flexible virtual fixtures to delimit safe work envelopes for shared human–excavator operation [8]. Subsequent studies introduced adaptive sliding mode regulators augmented with radial basis function (RBF) networks for electro-hydraulic servo loops [18] and an impedance controller that lets the bucket maintain a commanded contact force on variable soils [19]. Recent papers shift from single-cycle precision to mission-level efficiency: a multi-objective planner jointly minimizes time, energy, and ground impact for robotic digging [20]; an electro-hydraulic compensated servo enhances robustness under rock–soil transitions [21]; and a data-driven friction model based on the integration of GA and particle swarm optimization algorithm (PSO) was developed to obtain globally optimal solutions [22]. Collectively these studies outline a progression from component-level dynamics, through robust control, to safety-aware, energy-optimized fleet applications.

The green European cluster led by Marco Hutter and Philipp Leemann at ETH Zurich focuses on autonomous walking-excavator robotics: early work addressed legged-chassis force balancing and terrain-adaptive digging planners, followed by accurate free-form trenching, and culminating in the hydraulic excavator for an autonomous purpose (HEAP) platform, which integrates GNSS-RTK, multimodal sensing and dedicated drive, chassis, and arm controllers to enable fully autonomous multitask operation on uneven ground [23,24,25].

Despite an average link strength of 3.0, the overall structure remains sparse and regionally segmented: only four inter-cluster ties exceed a weight of two, and the network density is 0.18. The betweenness-centrality analysis identified Yin Chenbo (0.14) and Marco Hutter (0.12) as the main “knowledge brokers” capable of bridging Asian sensor-fusion research and European platform development. The persistence of strong intra-regional ties, coupled with weak cross-regional interaction, replicates patterns observed in earlier reviews of construction-robotics scholarship and underscores the field’s continuing fragmentation.

From an authorial perspective, three directions appear pivotal for accelerating knowledge diffusion. First, a community-curated, open benchmark repository—jointly overseen by the yellow and green clusters—would harmonize evaluation protocols across sensor-fusion and platform-design studies. Second, multinational pilot sites that pair European robotic trenching systems with Korean digital-twin workflows could create natural conduits for cross-cluster collaboration and technology transfer. Finally, structured mentorship or consortium schemes are needed to integrate peripheral researchers, especially those in adaptive hydraulics and building information modeling (BIM)-integrated simulation, into larger collaborative grants. Addressing these gaps would move the discipline beyond isolated regional breakthroughs toward globally standardized, interoperable solutions for autonomous earthmoving on complex construction sites.

4. Results and Discussion

As shown in

Table 2. Top 15 index keywords in the 157 selected papers.

Index Keyword	Frequency
Construction equipment	109
Excavation	91
Excavators	82
Automation	28
Robotics	28
Deep learning	21
Loaders	21
Construction industry	20
Hydraulic excavator	19
Wheels	18
Controllers	18
Autonomous excavators	17
Construction sites	17
Robotic excavator	16
Trajectories	16

, building on the bibliographic analysis, we first extracted the 15 most frequent index keywords from the 157-paper corpus, which reflect the field’s emphasis on heavy machinery, sensing, control, safety, and coordination. We then conducted open qualitative coding of abstracts, author keywords, and index keywords, iteratively grouping terms with similar meanings into four coherent research themes. Each theme is characterized by a collection of representative keywords that highlight its core focus.

By identifying and grouping recurring concepts such as sensor fusion, path planning, control algorithms, safety envelopes, and multi-machine coordination, we identified the following four major thematic dimensions of research:

• Autonomous Earthwork Machinery: This theme encompasses advances in environmental sensing and scene understanding. Representative keywords include deep learning; computer vision; simultaneous localization and mapping (SLAM); and light detection and ranging (LiDAR).
• Integrated Control Systems: Focusing on precise actuation and trajectory execution, this domain is reflected by keywords such as controllers; trajectories; sliding-mode control; model-predictive control; and adaptive control systems.
• Risk Mitigation Strategies: Covering methods to prevent, detect, and manage on-site hazards, this theme is signaled by keywords like collision-free trajectory planning; construction safety; active safety systems; and subsurface hazard detection.
• Fleet Interoperability: Addressing coordination among heterogeneous machines, this area is defined by terms such as unmanned aerial vehicles (UAV); unmanned ground vehicles (UGV); fleet operations; and multi-platform coordination.

In this section, we weave together the quantitative evidence from our keyword frequency analysis with the qualitative insights from our coding to explore each of these four pillars in depth, illustrating how they collectively span the sense–plan–act–collaborate cycle that underpins current research on autonomous earthwork machinery.

4.1. Autonomous Earthwork Machinery

The first and largest research cluster revolves around developing autonomous earthwork machinery [26]. Early surveys of construction-equipment automation highlighted key technical challenges and opportunities in this domain [27,28]. Chief among these challenges were enabling machine perception in unstructured environments and robustly controlling heavy hydraulic actuators. Over the past decade, rapid progress in sensing and AI has driven significant advances. For example, one recent study introduced a LiDAR-camera sensor-fusion framework that enhances an excavator’s terrain perception, allowing it to discern complex soil conditions and obstacles in real time [29]. Another work applied accelerated semantic segmentation and feature mapping to improve an excavator’s awareness of its surroundings, demonstrating reliable vision-based operation even in unstructured site conditions [30]. These developments exemplify the focus of the “automation” cluster identified earlier, with many publications dedicated to improving the autonomy of excavators, loaders, and related machinery.

Motion-planning capabilities have likewise matured [31]. Researchers have tailored path-planning algorithms to the dynamic and confined spaces of urban construction sites. A comparative study of global route planning in building environments showed the importance of accounting for vehicle-kinematic constraints when navigating cluttered sites [6]. At the vehicle level, Liu et al. proposed a hybrid A* algorithm with Dubins-curve smoothing for an articulated wheel loader, achieving an 18% reduction in path length during autonomous hauling operations [3]. Beyond navigation, the excavation process itself is being automated. A study in 2019 proposed a trajectory-planning method that integrates soil–bucket interaction physics and enables a hydraulic excavator to execute auto dig cycles with minimal human input [32].

Autonomy research now spans a variety of earthmoving equipment. For bulldozers, scholars have demonstrated complete-coverage path planning for automatic ground-levelling [33] and robust grading control under GPS-localization uncertainties [34]. Novel platforms have also emerged—most notably, the “HEAP” project realized an autonomous walking excavator capable of locomotion and digging on steep or uneven terrain [25]. Early milestones included a pressure-feedback hydraulic chassis that balances leg forces via quasi-static contact-force optimization [35] and an integrated planning control pipeline for autonomous digging on irregular ground [23]. Subsequent upgrades equipped the machine with excavation-specific 3D mapping and a limit-aware arm controller, enabling fully autonomous piecewise-planar and curved trenching with unprecedented accuracy along collision-free trajectories [24]. The latest HEAP prototype incorporates GNSS-RTK-referenced state estimation, multi-modal sensing and dedicated controllers for driving, chassis balancing and arm motion, and has been validated in field trials ranging from trench excavation and dry-stone-wall assembly to forestry manipulation and semi-autonomous teleoperation [25]. Together, these studies show that autonomous machine operation is a dominant theme in recent literature, driven by industry’s demand for higher productivity and reduced reliance on skilled operators.

4.2. Integrated Control Systems

A fully autonomous earthmover requires seamless coordination between perception, decision-making, and low-level actuation [36]. Advanced control-theory approaches help deliver the needed precision and robustness. Sliding-mode control (SMC) and its variants are widely used for their resilience to model uncertainty and external disturbances [18,37]. Fractional-order and adaptive SMC schemes for excavator arms have yielded improved tracking accuracy under varying soil resistance. Robust control has also been pursued: an H∞-based controller effectively rejected terrain-induced disturbances on a hydraulic manipulator, even when arm links flexed under heavy loads [38]. Model-predictive control (MPC) provides another avenue; an MPC implemented on a scaled excavator achieved precise motion tracking with zero steady-state error, demonstrating how predictive strategies handle actuator delays and constraints [39].

Control-integration research addresses the unique challenges of hydraulic drives. Accurate friction-modeling and compensation for excavator joints has significantly improved the smoothness of automated movements [22], while a time-delay control algorithm maintains responsiveness despite communication lags in tele-operated modes [40]. Machine-learning techniques are increasingly infused into control loops: deep reinforcement learning has been used to train a wheel-loader scooping controller, enabling autonomous optimization of bucket-filling motions [41]. Neural networks have likewise been blended with classical controllers to adapt to non-linearities in real time [18]. These integrated-control innovations bridge the gap between high-level autonomy and the rugged mechanical realities of earthmoving work.

4.3. Risk Mitigation Strategies

Safety remains a pivotal concern as autonomy moves from labs to live construction sites. Vision-based detection systems powered by deep learning can recognize construction personnel and other equipment in real time, alerting the autonomous machine—or halting it—to avoid dangerous proximity [42]. Coupled with sensor-fusion and tracking algorithms, such systems enable predictive safety evaluation, with the machine anticipating potential collisions along its path and adjusting proactively [7]. Hazard-specific sensing is also advancing: a magnetic system that detects buried metallic utilities allows an excavator to avoid subsurface obstacles before contact [43].

Operational strategies further reduce human exposure to danger. An earthwork digital-twin environment supported the teleoperation of an automated bulldozer during high-risk edge-dumping, keeping the operator safely off-site [44]. Fully autonomous solutions can remove workers entirely from hazardous zones; for example, an automatic excavation system used multiple synchronized excavators inside a pressurized pneumatic-caisson shaft, eliminating the need for personnel in the confined space [45]. These examples illustrate how autonomy, remote operation, and advanced sensing collectively underpin modern risk-mitigation practices.

4.4. Fleet Interoperability

Beyond single-machine autonomy, recent studies emphasize multiplatform cooperation—excavators, loaders, trucks, and even drones communicating and collaborating to accomplish tasks efficiently [1]. Industry roadmaps outline the capabilities required for connected fleets, including vehicle-to-vehicle interfaces and cloud-based site-management systems [46]. Practical progress is evident: deep-learning vision methods now track the real-time locations and actions of multiple construction machines simultaneously [47], enabling a central orchestration system to allocate tasks and prevent interference.

Digital platforms that coordinate fleet operations are emerging as well. An existing study developed a soil sharing system, which balances workloads and schedules excavation, hauling, and dumping across multiple machines [48]. Heterogeneous-agent collaboration is advancing vision-based unmanned aerial vehicle and unmanned ground vehicle (UAV–UGV) cooperation, which has shown drones mapping an area and guiding ground robots for site preparation [49]. Human–machine shared-control frameworks remain relevant; flexible virtual “fixtures” let a supervisor define safe zones that an autonomous excavator must respect [8], ensuring human insight and safety considerations are integrated even in highly automated fleets.

5. Research Gaps, Opportunities, and Priority Directions

The bibliometric and qualitative analyses in Section 3 and Section 4 revealed four thematic research domains, including autonomous machinery, integrated control, safety assurance, and fleet interoperability, each of which still have important technical and organizational gaps. To organize a forward-looking agenda, we systematically extracted and open-coded all gap statements, limitation remarks, and future work notes from the 157 papers’ abstracts, discussions, and conclusions. By clustering these gap codes according to their shared content (e.g., perception shortfalls, project integration issues, coordination barriers, safety certification needs, and human-automation challenges), we distilled the following five interrelated priority directions:

• Advanced perception and adaptive AI;
• Digital twins and integrated project management;
• Fleet collaboration and multi-machine autonomy;
• Safety assurance and regulatory frameworks;
• Human–automation collaboration and workforce adaptation.

In the subsections that follow, we examine each priority in turn—outlining its present limitations, identifying promising research pathways, and projecting its anticipated contributions to transforming isolated prototypes into dependable, self-optimizing fleets on urban construction sites.

5.1. Advanced Perception and Adaptive AI

Autonomous earthmoving machines rely on robust perception and intelligent decision-making to operate in unstructured, dynamic environments [36]. Recent efforts have focused on enhancing machine perception through advanced sensors and algorithms—for example, integrating multi-modal terrain sensing to improve excavators’ situational awareness [29] and developing SLAM techniques tailored to wheel loaders for richer site mapping [50]. Unstructured outdoor conditions with variable terrain and lighting demand perception methods that can interpret complex scenes in real time. Researchers have accelerated vision-based segmentation for terrain and object recognition in earthmoving contexts [30], and they have employed novel sensing modalities to tackle unique challenges, such as using hyperspectral imaging to discriminate material types like ore grade [51] or leveraging excavator force feedback to classify soil properties during digging [52].

Beyond perception, adaptive AI and control algorithms are being explored to handle the variability of earthmoving tasks. Reinforcement learning has shown promise for optimizing autonomous operations—for instance, improving a loader’s path efficiency and energy use [53] or learning complex excavator digging strategies through trial and error [41]. Likewise, data-driven approaches that leverage expert demonstrations (imitation learning) combined with optimization enable excavators to learn efficient trajectories for tasks like trenching and backfilling [54,55]. To support these AI-driven methods, large-scale datasets and simulation tools are increasingly utilized; for example, a new 3D point cloud dataset of heavy equipment has been used to train recognition algorithms [56], and synthetic imagery is being generated to augment training data for vision-based detection models [57]. These advances in perception and AI lay the groundwork for more autonomous and adaptive earthmoving, though achieving fully robust, real-time adaptation under all field conditions remains an ongoing challenge.

5.2. Digital Twins and Integrated Project Management

Industry roadmaps emphasize that connectivity and digital integration will be central to future construction automation [46]. A key concept is the use of digital twins which are live virtual replicas of the jobsite and machinery to integrate real-time field data with project planning models. Such approaches are gaining traction in earthmoving: by sharing soil and equipment data through cloud platforms, large-scale earthworks can be managed in a more transparent and optimized way [48]. Early prototypes demonstrate the value of linking autonomous machines to digital site models; for example, an earthwork “digital twin” system has been used to remotely supervise an automated bulldozer via real-time virtual feedback [44]. Similarly, continuous site sensing feeds dynamic project visualizations [58], enabling managers to view progress and equipment status in real time. Open data standards (e.g., open BIM for infrastructure) are being applied to connect autonomous machinery with project information models, facilitating seamless data exchange for monitoring and control [59,60]. This integration of field and office data promises to break down silos between equipment operation and project management. With these integrated systems, project managers can achieve enhanced oversight and decision-support. Automated tools now continuously monitor equipment productivity and progress using IoT devices and computer vision techniques. For example, multi-camera systems have been deployed to track earthmoving operations and calculate productivity metrics in real time [61], while drone photogrammetry combined with video analysis can rapidly assess earthwork progress and volume changes [62]. In parallel, researchers are automating complex planning tasks. Methods have been developed to automatically generate earthwork schedules and work breakdown structures using AI planning models [63,64]. Optimized scheduling of multiple machines is another active area, with genetic algorithms used to coordinate fleets of automated equipment for maximum efficiency [65,66]. Data-driven forecasting models are also leveraging field data to predict project outcomes: one study uses GPS tracked productivity data to forecast earthmoving costs and durations [67], and an expert system has been shown to optimize the equipment allocation and schedule for large earthwork projects [68]. Despite these innovations, fully integrating autonomous operations into traditional project workflows remains challenging. Issues of interoperability and data reliability persist, and user adoption is gradual [69,70]. Continued development of digital twin platforms and standardized data interfaces is needed to realize the full potential of an integrated, AI-driven project management ecosystem.

5.3. Fleet Collaboration and Multi-Machine Autonomy

Earthmoving automation is expanding from single-machine autonomy toward coordinated fleet operations. Recent reviews note that while this sector has pioneered automation, effectively orchestrating multiple machines (fleet automation) is a frontier for improving productivity [27]. Multi-machine autonomy offers opportunities for heavy equipment to work in concert. For instance, the synchronized control of multiple excavators in specialized foundation construction methods [45] or cooperative teams of aerial drones and ground vehicles for site preparation tasks [49]. To orchestrate such collaboration, reliable inter-machine communication and coordinated planning are essential. Research has demonstrated control frameworks for excavator clusters operating together on shared tasks [11], and “Internet of Vehicles” architectures have been proposed to network smart machines and enable real-time data exchange among equipment on site [9,71]. Multi-agent task allocation algorithms are now being developed to dynamically distribute work among robotic fleets, aiming to optimize overall throughput and minimize idle time [1]. For example, field studies using vision-based analysis of excavator–truck interactions can identify cycle bottlenecks, informing better coordination strategies between loading and hauling equipment [72]. Enhanced situational awareness is also crucial in these scenarios; recent methods use video and sensor fusion to track the positions and movements of multiple machines simultaneously, helping to prevent conflicts in shared work areas. While early demonstrations of cooperative earthmoving are promising, achieving seamless fleet-wide autonomy remains challenging. Heterogeneous machines must manage complex spatial interactions, communication lags, and dynamic task re-prioritization, all while upholding safety in proximity operations [47]. Addressing these issues will require further advances in multi-robot planning, robust wireless networks, and fleet management frameworks tailored to construction environments.

5.4. Safety Assurance and Regulatory Frameworks

As earthmoving machinery becomes more autonomous, ensuring operational safety and establishing rigorous validation frameworks have become paramount [28]. Active research is addressing these concerns: for example, advanced sensor-fusion algorithms now track nearby objects and predict potential collisions in real time, allowing an autonomous excavator to preemptively avoid accidents [7]. Vision-based safety systems can detect personnel or other equipment in the vicinity of a machine and trigger automatic interventions, effectively creating a virtual safety bubble around working robots [73]. In addition to visible hazards, novel sensing methods are tackling hidden dangers underground—magnetometer arrays can locate buried metal pipes to prevent accidental strikes during digging [74], and related avoidance systems aim to protect subsurface infrastructure by warning or halting the machine when underground utilities are detected [43]. These technological measures enhance on-site safety as automation increases. Developing a formal regulatory framework for autonomous construction equipment is an equally important challenge. Unlike highway vehicles, off-road construction robots currently lack well-defined safety standards or certification processes. Industry stakeholders and researchers recognize the need for clear guidelines to govern the deployment of autonomous heavy equipment [28]. Ensuring that autonomous systems can demonstrate at least equivalent safety to human operators is critical for gaining public and regulatory acceptance [26]. A useful reference is ISO 17757:2019, which specifies functional-safety requirements for autonomous and semi-autonomous earth-moving machinery and mining [75]. Although written for open-pit contexts, its risk-assessment framework, machine-stop categories, and human–machine-interface provisions can be transposed to urban construction with only moderate adaptation. For example, adding rules for mixed traffic and tighter geofence precision. Pilot projects apply ISO 17757 checklists during factory acceptance tests of tele-remote dozers and autonomous haulers, indicating its practical value while a construction-specific standard is still under development [76]. To this end, there are calls for standardized testing protocols (e.g., scenario-based simulations and field trials) and transparent safety assessment methodologies for construction robots [28]. Close collaboration between equipment manufacturers, construction firms, and regulators will be required to develop policies that keep pace with technological advancements. Establishing consensus on liability and compliance standards will further pave the way for the safe and widespread adoption of autonomous earthmoving machinery.

5.5. Human–Automation Collaboration and Workforce Adaptation

Even with growing automation, human operators and engineers remain central to the earthmoving process. Teleoperation technology, for instance, enables people to control machines from a safe distance or centralized control room, and has been successfully tested in demanding contexts like simulated lunar construction with time delays [77]. Such remote and assisted operation can extend human capabilities but also highlights the need for intuitive interfaces to maintain performance under challenging conditions (e.g., handling communication lags). Rather than replacing operators, current research often aims to augment their performance. Semi-autonomous assistive control systems can handle repetitive low-level motions or stabilize the machine, allowing the human to focus on high-level decisions [8,10]. In a shared-control paradigm, the human and AI system split responsibilities—for example, an algorithm might identify optimal digging targets or adjust the excavator’s motion path in real time while the operator oversees the general task [78]. New human–machine interaction modalities are also being explored to make control more intuitive; a recent study demonstrated gaze-aware hand gesture recognition for commanding construction machines, enabling an operator to give natural, contactless instructions to an intelligent excavator [79]. A significant future challenge lies in preparing the workforce to effectively collaborate with these automated systems. The introduction of robotics on sites has been slow in part due to human factors and organizational concerns. Gaining operator acceptance and trust in automation is as important as the technology itself. Training programs and upskilling initiatives will be needed to help equipment operators transition into roles that combine supervision and technical oversight of autonomous fleets. In the near term, a human–automation collaborative approach is envisioned: human experts will increasingly act as supervisors or mission managers for multiple semi-autonomous machines, rather than manually controlling a single machine at all times. Studies charting the path from tele-remote operation to semi-autonomy underscore that operators can gradually shift from direct control to a monitoring and exception-handling role as the technology matures [80]. Ultimately, researchers even imagine largely unsupervised robotic construction sites in the future [81], but realizing this vision will require not only technical innovation but also careful integration of human insight. The knowledge and experience of veteran operators must be captured and reflected in autonomous systems, and the workforce must be intimately involved in the technology’s development and deployment. By proactively addressing human–automation collaboration and workforce adaptation, the construction industry can ensure that automation enhances productivity and safety without disenfranchising its skilled workers.

6. Conclusions

This review set out to clarify how research on autonomous earth-work machinery has evolved over the past decade and what remains to be done before fully automated fleets can operate routinely on urban construction sites. By screening 597 Scopus records and analyzing 157 relevant articles with a mixed bibliometric-and-content approach, we identified four inter-dependent research strands—machine autonomy, integrated control, safety assurance, and fleet coordination—that together define the state of the art.

The literature makes it clear that widespread deployment will depend on closing several concrete gaps. Perception pipelines that falter under dust, rain, or glare must be hardened through self-supervised adaptation and multi-modal redundancy; robust, low-latency communication stacks are still needed to sustain cross-vendor collaboration among heterogeneous fleets; and the field lacks scenario-based validation protocols comparable to ISO 26262 in automotive engineering, leaving regulators with little basis for certifying functional safety. Equally pressing is the integration of autonomous machinery with BIM-centric workflows: without standardized data interfaces, digital twins cannot provide the continuous feedback loops on which truly self-optimizing fleets rely. Finally, a sustainable human–automation partnership has yet to be articulated—operators must evolve from manual control to mission supervision, but curricula and interface designs that support this transition remain embryonic. Beyond the technical hurdles, the sector must also confront socio-technical tensions: autonomous fleets may displace certain machine-operator roles while creating demand for higher-skilled system supervisors and maintenance engineers; proactive retraining schemes and inclusive technology-adoption policies will be essential to ensure that productivity gains do not come at the expense of workforce wellbeing and community acceptance.

In sum, the field has advanced from concept to pilot, but the transition to scalable, dependable practice will hinge on addressing the technical and organizational gaps identified here. The roadmap distilled in this review—adaptive perception, digital-twin integration, cooperative autonomy, certified safety, and human–automation partnership—provides a focused agenda for the next generation of research and development in autonomous urban earthmoving.