When Robots Learn: A Bibliometric Review of Artificial Intelligence in Engineering Applications of Robotics

Abstract

The convergence of Robotics and artificial intelligence (AI) has transformed engineering by enabling the design of intelligent systems capable of learning, adapting, and performing complex tasks. These synergies are driving innovation across multiple engineering disciplines, including mechanical, materials, electrical, industrial, civil, and aerospace engineering. This review provides a comprehensive overview of the knowledge structure and emerging research directions of Robotics and AI in engineering, with the aim of identifying research trends, influential authors, leading institutions, and emerging thematic areas. Data were collected from the Web of Science and Scopus databases, covering the period from 2020 to 2025, and analyzed using bibliometric mapping techniques and performance indicators. The results reveal a sustained growth in research on autonomous systems, collaborative robots, and human–robot interaction within engineering contexts, with a strong emphasis on AI-driven optimization. Bibliometric analyses show that deep learning, reinforcement learning, and computer vision constitute the core enabling technologies structuring the field. In addition, the results highlight a high degree of international collaboration and a concentration of scientific output and impact in a limited number of leading countries, institutions, and journals.

1. Introduction

The integration of Artificial Intelligence (AI) into robotic systems has represented a profound transformation in the field of engineering over the last decade [1]. The combination of perception, reasoning, and automatic learning capabilities with mechanical and electronic systems has enabled the development of robots capable of adapting to complex environments, interacting with humans, and performing tasks with a high degree of autonomy [2]. This technological convergence is driving significant advances in multiple branches of engineering, such as mechanical engineering [3], electrical, industrial, and civil engineering [4], materials engineering [5] and aerospace engineering [6], consolidating intelligent robotics as one of the pillars of contemporary engineering [7].

Although the present study adopts a transversal bibliometric perspective, the analyzed literature spans differentiated engineering domains—including manufacturing robotics, infrastructure inspection in civil engineering, aerospace autonomous systems, and advanced mechanical automation—each characterized by specific operational constraints and application contexts. This cross-domain scope reflects the systemic diffusion of AI-driven robotics across engineering disciplines rather than a focus on a single sector.

Scientific interest in AI-based robotics has grown significantly, driven both by the maturity of techniques such as deep learning and reinforcement learning [8,9] and by the need for efficient solutions in industrial, logistics, healthcare, and infrastructure contexts. In this scenario, autonomous systems, collaborative robots, particularly in relation to trust calibration and safety assessment in human–robot collaboration [10] and human–robot interaction [11], have become priority research lines, reflecting a paradigm shift from rigidly programmed robots toward adaptive systems oriented toward intelligent decision-making.

Despite the high volume of published research, existing knowledge remains fragmented across different disciplines and methodological approaches [12]. Most studies focused on specific technical developments or narrative reviews of particular technologies, which makes it difficult to obtain a global and structured view of the field [13]. In addition, relevant debates persist, such as the balance between autonomy and human supervision, the reliability of learning algorithms in critical environments, or the transferability of AI-based solutions across different engineering domains [14,15]. These divergences highlight the need for systematic analyses that allow for the identification of patterns, consensuses, and research gaps. In this context, bibliometric studies are presented as an appropriate tool to systematically and objectively analyze the evolution of a scientific field characterized by high interdisciplinarity [16]. Through the use of performance indicators and scientific mapping techniques, bibliometrics makes it possible to identify temporal trends, key actors, collaboration networks, and underlying conceptual structures [17,18].

This interdisciplinary complexity is especially evident in the convergence between robotics and artificial intelligence, where dimensions such as perception, machine learning, autonomy, and human–robot interaction act jointly to enhance engineering capabilities, as schematically illustrated in Figure 1 [11]. This integrative approach highlights that progress in the field cannot be analyzed from a single discipline, but rather as the result of interactions among multiple technological and conceptual areas.

However, despite the growing number of publications on robotics and artificial intelligence, bibliometric analyses that specifically and transversally address their application in the field of engineering remain scarce, considering jointly their impact in different engineering branches [19]. This shortcoming justifies the need for a study that offers an integrated view of the field and contributes to clarifying its structure, dynamics, and emerging lines of development [16].

In light of this scenario, this bibliometric study constitutes an appropriate tool to objectively analyze the evolution of the convergence of robotics and Artificial Intelligence to transform engineering. The main objective of this work is to carry out an exhaustive bibliometric review of the scientific literature on robotics and artificial intelligence applied to engineering. Through the analysis of scientific production, collaboration networks, and the thematic structure of the field, this study aims to offer an integrated view of the current state of research, identify the areas of greatest impact, and point out the main lines of future development [16].

2. Materials and Methods

The present study adopted a bibliometric methodology with the aim of systematically analyzing the evolution, intellectual structure, and main trends [20] of scientific research related to the application of Artificial Intelligence in Robotics within the field of engineering. This approach makes it possible to evaluate large volumes of scientific literature through quantitative indicators and mapping techniques, providing an objective and reproducible view of the development of the field.

Bibliographic data were collected from the Web of Science (WoS) and Scopus databases, selected for their broad multidisciplinary coverage, the quality of indexed records, and their established use in bibliometric studies. The combined use of both sources makes it possible to minimize biases derived from indexing and to obtain a more comprehensive representation of relevant scientific production [21].

The search strategy was carefully designed to capture specific scientific literature addressing the convergence between robotics, artificial intelligence, and engineering applications. The selection of terms included in the search equation was based on keywords widely used in scientific literature, as well as on terms representative of the main subfields of artificial intelligence and robotic systems [20].

As shown in

Table 1. Search String. The wildcard (*) indicates term variations.

Search String
(“robot*” OR “autonomous system*”) AND (“artificial intelligence” OR “AI” OR “machine learning” OR “deep learning” OR “neural network*” OR “intelligent system*”) AND “engine*”

, the search string was designed to identify academic or technical literature at the intersection of robotics [22], autonomous systems, and artificial intelligence [23] applied to the field of engineering [24]. Similarly, it sought to identify engineering contexts, as well as aspects related to technological innovation or manufacturing. This search string was applied to the title, abstract, and keywords fields in both databases, using Boolean operators and truncations to maximize the retrieval of relevant documents.

The selected time period spanned from 2020 to 2025 (both inclusive), with the aim of analyzing the most recent evolution of the field of robotics and artificial intelligence applied to engineering [25]. This interval coincides with a phase of strong scientific and technological expansion, characterized by the consolidation of advanced machine learning techniques and their progressive adoption in engineering applications [26] and in project management [27]. Likewise, the choice of this period makes it possible to capture current trends in the area, avoiding the inclusion of earlier works that correspond to approaches already surpassed or to initial stages of technological development [28].

The process of identification, selection, and screening of documents was carried out following the PRISMA 2020 protocol (Figure 2), in order to ensure the transparency, reproducibility, and methodological rigor of the literature selection process [29].

In the identification phase, a total of 2435 records were retrieved from the Web of Science database and 2791 records from Scopus using the defined search string, resulting in 5226 initial records. After combining both datasets, 1472 duplicate records were removed, yielding 3754 unique documents.

Subsequently, in the screening phase, a first exclusion criterion (EC 1) based on language was applied, excluding documents not written in English, and retaining only those written in English [30]. As a result of this filtering, the dataset was reduced to 2400 records. Subsequently, an exclusion criterion (EC 2) related to document type was established, removing those records that did not correspond to peer-reviewed scientific publications. In a complementary manner, a first inclusion criterion (IC 1) was defined, through which only article and review-type documents were selected, as these formats provide original research or consolidated syntheses of knowledge, while other document types were excluded, such as conference proceedings, book chapters, editorials, letters, corrections, and meeting abstracts, due to their preliminary nature or lower bibliometric impact. As a result of this filtering, 739 records were excluded.

Finally, in the inclusion phase, the definitive set of documents subject to analysis consisted of 1661 publications, which formed the basis of the bibliometric analysis carried out in this study [31,32].

In order to provide a clear and structured overview of the results obtained, as can be observed in Figure 3, the bibliometric analysis was organized into five main categories [33], widely used in previous bibliometric studies: (i) general information of the study, (ii) scientific sources, (iii) authors and collaboration patterns, (iv) scientific impact, and (v) conceptual structure of the research field [34].

The bibliometric analysis was conducted using the Bibliometrix package (version 5.2.1; University of Naples Federico II, Naples, Italy; https://www.bibliometrix.org), executed in the RStudio environment (version 4.5.1; Posit Software, PBC, Boston, MA, USA; https://posit.co). This package made it possible to perform both performance analysis and scientific mapping, including co-authorship, co-citation, and keyword co-occurrence networks.

The selection of bibliometric indicators followed established guidelines for science mapping studies, combining performance indicators (publication output, citation impact, collaboration patterns) with relational mapping techniques (co-authorship, co-citation, and keyword co-occurrence) in order to capture both quantitative productivity and conceptual structure. Multiple correspondence analysis (MCA) was employed for clustering and factorial mapping because it is particularly suitable for analyzing keyword co-occurrence matrices and identifying latent thematic structures in large bibliographic datasets. This methodological combination enabled a balanced interpretation of structural evolution, thematic organization, and knowledge diffusion within the field.

3. Results

Figure 4 presents a general summary of the main bibliometric indicators corresponding to the final set of 1661 documents published between 2020 and 2025. These indicators provide a quantitative characterization of the analyzed corpus and serve as a starting point for the detailed analysis of the different dimensions of the study. According to the data shown in Figure 4, the analyzed period comprised a total of 771 scientific sources, which highlights the high dispersion of scientific production across a wide range of journals and specialized publications [35].

The total number of identified authors amounted to 5468, with an average of 4.79 co-authors per document, which reflects a pronounced collaborative nature of research in this field [31]. Nevertheless, the presence of 114 single-author documents was also observed, indicating that, although minoritarian, individual research continues to play a relevant role [32].

Regarding document types, research articles clearly predominated (1347 documents), followed by reviews (314 documents), in line with the inclusion criteria adopted. This distribution confirms that the analysis was based mainly on peer-reviewed contributions that provide original results or consolidated syntheses of knowledge.

The analysis of keywords revealed a high number of terms both in author keywords (6099) and in Keywords Plus (2787), which suggests a notable thematic and terminological diversity within the field [36]. Likewise, the percentage of international co-authorship reached 30.04%, evidencing a significant degree of collaboration among researchers from different countries.

As observed in Figure 5, the annual evolution of the number of publications showed a clearly increasing trend throughout the 2020–2025 period. After a progressive increase between 2020 and 2023, a significant acceleration was observed from 2024 onwards, with the maximum volume of publications being reached in 2025. This pattern highlights the growing scientific interest and consolidation of artificial intelligence applied to robotics within the field of engineering.

The analysis of scientific sources made it possible to identify the journals that concentrated the highest number of publications related to the application of artificial intelligence in robotics within the field of engineering. Figure 6 shows the ten most productive journals in terms of the number of articles published during the analyzed period.

The results indicate that IEEE Access was the source with the highest number of publications, standing clearly above the other journals. This result reflects the multidisciplinary and broad-scope nature of this publication, as well as its high capacity to host research related to robotics, artificial intelligence, and engineering applications. In second place, the journal Sensors stood out, followed by Applied Sciences, both showing a notable presence of works focused on intelligent systems, sensing, and technological applications.

Likewise, a significant contribution from journals specialized in robotics and automation was observed, such as IEEE Robotics and Automation Letters and IEEE Transactions on Automation Science and Engineering, which highlights the relevance of these forums in the dissemination of advanced research on intelligent robotic systems. Other relevant sources included Engineering Applications of Artificial Intelligence, Electronics, Machines, and journals oriented toward systems and cybernetics, reflecting the disciplinary diversity of the field.

Figure 7 presents the cumulative temporal evolution of the number of publications in the main journals addressing robotics and artificial intelligence applied to engineering during the 2020–2025 period. All of the analyzed sources showed an increasing trend, reflecting the sustained growth of scientific production in this field. IEEE Access stood out as the journal with the highest cumulative growth, followed by Sensors and Applied Sciences, while IEEE Robotics and Automation Letters and IEEE Transactions on Automation Science and Engineering presented a more gradual but steady evolution. Overall, these results confirm the role of these journals as relevant channels for the dissemination of research in robotics and artificial intelligence applied to engineering.

The analysis of author productivity made it possible to identify the researchers with the highest number of publications in the field of robotics and artificial intelligence applied to engineering. In this context, it was observed that a small group of authors concentrated a significant number of articles published during the analyzed period. Among them, authors such as Y. Zhang and J. Li stood out, showing the highest values in terms of total scientific production, followed by Y. Wang, J. Zhang, and X. Wang.

When considering the number of fractionalized articles, a notable decrease was observed compared to the absolute count of publications. This behavior, illustrated in Figure 8, highlights the high participation of multiple co-authors in most of the analyzed works, which is characteristic of research areas with a strong interdisciplinary and technological component. The difference between both indicators therefore makes it possible to nuance the individual contribution of the most prolific authors.

Among the analyzed institutions, the University of California System showed the highest cumulative production of articles, with particularly marked growth from 2023 onwards. A sustained growth pattern was also observed at Tsinghua University, which maintained an upward trajectory throughout the entire period considered. Other relevant institutions included Sun Yat-sen University, Shanghai Jiao Tong University, and the Chinese Academy of Sciences, all of which showed progressive increases in the number of publications.

The cumulative temporal evolution of these institutions, represented in Figure 9, highlights a generalized increase in research activity, especially in the last years of the analyzed period. Although the trajectories showed different growth rates, all institutions displayed a clear upward trend, reflecting their continued involvement in the development of research related to the integration of artificial intelligence into robotic systems applied to engineering.

The analysis of scientific production by countries, considering the corresponding author’s country, made it possible to identify both the relative weight of each nation in the total volume of publications and their international collaboration patterns. As shown in Figure 10, China clearly led scientific production in the field of robotics and artificial intelligence applied to engineering, with 506 documents, representing 30.5% of the total. Of these, 387 corresponded to single-country publications (SCPs), while 119 were developed through international collaboration (MCP), reflecting a combination of strong internal research capacity and a relevant presence in global collaboration networks.

The United States occupied the second position, with 230 articles (13.8%), also showing a significant balance between national publications (179 SCPs) and international ones (51 MCPs). Together, these two countries concentrated a substantial part of the scientific production of the field, consolidating themselves as the main research hubs at the global level.

A second group of countries, led by India, the United Kingdom, South Korea, Italy, and Germany, showed a more moderate contribution in absolute terms, but with differentiated patterns of international collaboration. The cases of the United Kingdom, Italy, Canada, and Australia stood out in particular, where the percentage of internationally co-authored publications exceeded 40%, reaching values above 50% in countries such as Canada and Australia. This behavior suggests greater integration into international scientific networks and a research strategy strongly oriented toward transnational cooperation.

Overall, these results show that research in robotics and artificial intelligence applied to engineering presents a marked global dimension, in which leading countries combine high national production with increasing participation in international collaborations, contributing to the advancement and dissemination of knowledge in the field.

The analysis of scientific impact, measured through the total number of citations by country, reinforces these observations. As shown in Figure 11, China and the United States clearly led the ranking of the most cited countries, with a notable difference compared to the rest. At a greater distance were the United Kingdom, India, and Australia, while European countries such as Italy, Germany, the Netherlands, and France presented more moderate citation levels, although still relevant within the global context.

Table 2. Most global cited documents.

Ref.	Authors	Scientific Journal	Total Citations	TC per Year
[37]	C. Li	IEEE Transactions on Image Processing	1469	209
[38]	H.Y. Pan	Automatica	655	109
[39]	L. Wright	Nature	503	100
[40]	M. Lezoche	Computers in Industry	499	71
[41]	D. Christensen	Neuromorphic Computing and Engineering	463	92
[42]	Z. Wang	Reports on Progress in Physics	450	75
[43]	S. Kelly	Telecommunication Systems and Information Processing	446	111
[44]	M. Cao	Advanced Materials	410	58
[45]	V. Lu	Journal of Service Theory and Practice	404	57
[46]	A. Darko	Automatica	378	54

lists the most cited articles within the set of analyzed documents, highlighting those works that have exerted the greatest scientific impact in the field of robotics and artificial intelligence applied to engineering. The results showed that some articles concentrated a very high number of total citations, especially those published in high-impact journals such as IEEE Transactions on Image Processing, Automatica, and Nature. Likewise, the citations-per-year indicator made it possible to observe differences in the rate of impact accumulation, reflecting both the sustained relevance of older works and the rapid influence of more recent publications. Overall, these results highlight the existence of key contributions that have acted as fundamental references in the development of the field during the analyzed period.

Figure 12 presents a tree-map of the most frequent keywords, offering a global view of the thematic structure of the field. The results show that AI constituted the dominant term, concentrating the highest percentage of occurrences, followed by machine learning, deep learning, and robots, which confirms the central role of machine learning techniques in current research on robotics applied to engineering. Likewise, terms related to system design and modeling, such as design, model, and system, appeared with relevance, reflecting the engineering-oriented focus of the field.

At a second level, specific topics such as optimization, automation, sensors, and reinforcement learning were identified, which act as enabling technologies for the development of advanced robotic systems. Finally, the presence of emerging terms linked to human–robot interaction, computer vision, navigation, and feature extraction evidenced a thematic diversification oriented toward more autonomous, adaptive, and user- and environment-centered applications.

4. Discussion

The analysis of the conceptual structure of the field revealed the existence of three well-defined thematic clusters, whose distribution and relationships are represented in Figure 13. These clusters reflect the main research lines that articulate the integration of artificial intelligence into robotics applied to engineering, organized along the main dimension, which separates methodological and conceptual approaches from those clearly oriented toward practical application.

In this factorial map, dimensions Dim1 and Dim2 correspond to the first two axes obtained from a multiple correspondence analysis applied to the keyword co-occurrence matrix. These dimensions represent the directions of greatest variability in the conceptual structure of the field and allow the terms to be projected into a two-dimensional space that preserves, as far as possible, their semantic relationships.

Dimension 1 (Dim1) can be interpreted as an axis that separates methodological and algorithmic approaches, associated with the foundations of artificial intelligence and machine learning, from those oriented toward the application and practical deployment of robotic systems in engineering contexts. In turn, Dimension 2 (Dim2) distinguishes between works focused on perception and information acquisition processes, such as sensing and feature extraction, and those focused on decision-making processes, interaction, and system performance.

The relative position of the terms in this space reflects their degree of conceptual association, such that closely positioned terms tended to co-occur more frequently in the literature, while the identified clusters emerged as coherent groupings of topics that share a common conceptual basis.

Cluster I was mainly associated with the most applied and operational aspects of robotic systems and was clearly located at positive values of the main dimension. This group concentrated terms such as computational modeling (3.00), training (2.18), task analysis (2.14), robot sensing systems (1.73), accuracy (1.41), and robots (1.46). This location indicates a marked orientation toward practical implementation, system training, and performance evaluation in real engineering contexts. The cluster reflects the interest in transferring advances in artificial intelligence to functional robotic solutions, where accuracy, reliability, and performance optimization constitute central objectives.

Cluster II, smaller in size but conceptually well-defined, grouped terms related to perception and the initial processing of information and was located in an intermediate area of the main dimension, with moderately positive values. Concepts such as sensors (0.33) and feature extraction (0.83) highlight the fundamental role of sensing and data processing as enabling elements of intelligent robotic systems. Their intermediate position suggests that this cluster acts as a bridge between methodological developments and practical applications, facilitating the connection between artificial intelligence algorithms and their effective deployment in real robotic systems.

Finally, Cluster III encompassed the fundamental concepts of artificial intelligence and machine learning and was predominantly located at negative values of the main dimension. This group included terms such as artificial intelligence (−0.52), machine learning (−0.29), deep learning (−0.09), neural networks (−0.66), reinforcement learning (0.32), and human–robot interaction (−0.10). This distribution indicates a more theoretical and methodological orientation, focused on the development of algorithms, learning models, and conceptual frameworks that endow robots with adaptive, cognitive, and advanced interaction capabilities.

The relevance of these clusters is further supported by the thematic coherence observed in highly cited publications within the dataset. Applied and operational research (Cluster I) is typically reflected in studies focused on industrial implementation, system optimization, and performance validation in real engineering contexts. Perception-oriented research (Cluster II) is commonly associated with contributions centered on sensing technologies, computer vision, and feature extraction mechanisms. Likewise, the methodological core (Cluster III) corresponds to foundational works addressing deep learning architectures, reinforcement learning strategies, and intelligent modeling frameworks. This alignment between cluster structure and dominant publication themes reinforces the internal consistency and validity of the identified thematic organization.

The analysis of the thematic map makes it possible to further examine the degree of development and relevance of the main themes that structure research in robotics and artificial intelligence applied to engineering, as observed in Figure 14. The distribution of themes according to their centrality and density provides a clear view of their state of maturity and the role they play within the field.

Motor themes were located in the upper-right quadrant and grouped concepts such as artificial intelligence, robotics, and automation, together with design and optimization. The position of these themes indicates that they constitute well-developed and highly relevant cores, acting as structuring axes of current research. Their high centrality reflects their connection with multiple lines of work, while their density suggests a high degree of conceptual and methodological consolidation.

In the lower-right quadrant, basic themes were located, among which robots, reinforcement learning, and training stood out. These themes showed high relevance for the field as a whole, but a lower degree of internal development, indicating that they function as widely shared foundations upon which more specific research is articulated. Their position suggests that, although they are essential, they still offer room for greater deepening and specialization.

Emerging or declining themes, located in the lower-left quadrant, included concepts such as machine learning, deep learning, models, and architecture. This location can be interpreted as an indication of thematic transition: on the one hand, some of these concepts have reached a degree of maturity that shifts them toward other, more applied clusters; on the other hand, it suggests the emergence of new approaches or methodological reformulations within the field.

The overall set of results obtained highlights a clear evolution of the field from fundamentally methodological approaches toward robotic applications fully integrated into engineering contexts. The identified conceptual structure, together with the temporal distribution of publications and the hierarchy of keywords, suggests that research in artificial intelligence applied to robotics has moved beyond an initial phase focused on the development of generic algorithms toward solutions oriented to the implementation, training, and optimization of real robotic systems. This shift reflects a progressive maturation of the field, in which theoretical advances in machine learning are increasingly translated into functional applications with a direct impact on different engineering domains.

From a theoretical standpoint, this evolution reflects a broader paradigm shift in engineering research, where artificial intelligence is no longer conceived as an isolated computational tool but as an embedded component of cyber-physical systems. The convergence between learning algorithms, sensing technologies, and domain-specific engineering constraints indicates a transition toward system-level integration models, in which performance, safety, and adaptability must be jointly optimized. This perspective moves the discussion beyond descriptive mapping and highlights the conceptual implications of AI-driven robotics for engineering theory and practice.

Beyond descriptive bibliometric mapping, the present study contributes by identifying structural patterns in the evolution of AI-driven robotics within engineering, clarifying the transition from algorithm-centered research toward deployment-oriented applications. The integration of conceptual clustering, thematic evolution, and country-level impact analysis provides a systemic interpretation of how methodological advances are progressively translated into engineering practice. In this sense, the study offers not only quantitative indicators but also a structured analytical framework for understanding the maturation of the field.

In this sense, the increasing relevance of terms associated with training, computational modeling, accuracy, and robot sensing systems indicates that the focus of research is shifting toward improving the performance, reliability, and robustness of intelligent robotic systems. This trend suggests a consolidated interest in validating and deploying artificial intelligence-based solutions in complex and operational environments, where engineering-specific constraints—such as safety, efficiency, or reproducibility—play a central role. In a complementary manner, the persistence of core concepts such as machine learning, deep learning, and neural networks confirms that methodological development remains an essential pillar, although it is now clearly oriented toward supporting practical applications and facilitating the transfer of knowledge from the laboratory to real engineering scenarios [47,48].

Beyond methodological advances and final applications, the results of the bibliometric analysis highlight the key role of a set of enabling technologies that act as a link between artificial intelligence algorithms and their effective implementation in engineering robotic systems. In particular, the recurrence of terms such as sensors, feature extraction, computer vision, and modeling shows that perception, information processing, and environment representation constitute essential elements for translating machine learning capabilities into functional robotic behaviors.

These technologies play a transversal role in the field, facilitating the acquisition of reliable data, environment interpretation, and real-time decision-making. Advanced sensing and feature extraction make it possible to transform complex physical signals into structured information that can be exploited by learning algorithms, while computer vision techniques expand robots’ ability to interact autonomously with dynamic and unstructured environments. Complementarily, computational modeling and simulation are consolidated as fundamental tools for the design, training, and validation of intelligent robotic systems, reducing costs, risks, and development times [49].

The intermediate position occupied by these technologies in the conceptual structure of the field suggests that they act as an integrating element, connecting theoretical developments in artificial intelligence with scalable multi-robot implementations in engineering contexts [50]. This bridging function is key to the scalability and transferability of AI-based robotic solutions, enabling their adaptation to different engineering domains and applications and reinforcing the interdisciplinary and applied nature of current research.

Despite the notable progress and the progressive maturation of research on artificial intelligence applied to robotics in engineering, the results of the present study reveal the persistence of several bottlenecks that limit the effective integration of these technologies in complex engineering environments. One of the main challenges identified is related to the robustness and reliability of intelligent robotic systems, especially when they operate in dynamic, uncertain, or partially structured scenarios. The dependence on large volumes of data and the sensitivity of learning algorithms to variations in the environment limit, in many cases, the direct transfer of solutions developed in controlled settings to industrial or field applications [51]. These limitations become particularly critical in highly regulated or safety-critical engineering domains, where performance degradation may compromise operational reliability.

In particular, the quality of computer vision datasets, the precision of annotation, and discrepancies between training and deployment environments significantly influence model performance. Noise, labeling inconsistencies, domain shifts, lighting variations, sensor degradation, and occlusions may lead to error propagation in perception pipelines, especially in safety-critical engineering applications. These factors reinforce the need for robust dataset design, domain adaptation strategies, and continuous validation under real operating conditions.

Another critical aspect is the safety and verifiability of artificial intelligence-based systems. In engineering applications where robots interact with people, infrastructures, or critical processes, it is essential to ensure predictable and controllable behaviors. In this context, this means research lines oriented toward model explainability, formal validation of algorithms, and the integration of human supervision mechanisms that make it possible to balance autonomy and control gain relevance [52]. This represents a structural bottleneck for large-scale industrial deployment, where compliance with engineering standards and predictable behavior are mandatory requirements.

Likewise, the transferability and scalability of AI-based robotic solutions emerge as a key challenge. The results suggest that many proposals remain strongly tied to specific domains or tasks, which hinders their reuse in different engineering contexts. In this regard, the development of modular architectures, generalizable models [53], and learning strategies that reduce dependence on labeled data constitute a priority line of work for future research. Overall, robustness limitations, safety constraints, transferability issues, and data dependency emerge as the main bottlenecks hindering the scalable integration of AI-enabled robotic systems in complex engineering contexts. Future research should therefore prioritize the development of hybrid approaches combining data-driven learning with model-based control, improved sim-to-real transfer methodologies, explainable AI frameworks, and standardized validation protocols. Strengthening interdisciplinary collaboration between AI researchers, control engineers, and safety specialists will be essential to ensure the deployment of more robust and trustworthy autonomous robotic systems.

As a global synthesis, Figure 15 makes it possible to integrate the main trends, challenges, and future directions of the area, providing a structured view of the current state and the evolutionary potential of intelligent robotics in engineering. The growing incorporation of intelligent robots into productive and social environments raises ethical and organizational challenges that go beyond the purely technical domain. Aspects such as human–robot interaction, social acceptance, responsibility in automated decision-making, and the impact on work processes require multidisciplinary approaches that integrate technical, human, and regulatory considerations. Regarding collaborative robots, current research on human–robot interaction increasingly addresses both (i) physical safety (collision avoidance, safe motion planning, speed-and-separation monitoring, and fail-safe behaviors) and (ii) psychological safety (trust calibration, perceived safety, workload, and user acceptance). While physical safety has benefited from more mature sensing and control approaches, psychological safety boundaries remain more context-dependent and less standardized, particularly in complex engineering environments where humans and robots share dynamic workspaces. As a result, a growing body of work emphasizes human-centered evaluation protocols and the integration of transparency and explainability mechanisms to improve predictability and perceived safety during collaboration. Addressing these challenges jointly will be decisive in consolidating the effective and sustainable deployment of intelligent robotics in the engineering of the future.

Overall, the results of the bibliometric analysis show that research on artificial intelligence applied to robotics in engineering is in a phase of consolidation, characterized by the convergence of methodological developments, enabling technologies, and engineering applications. The evolution of the field reflects a progressive shift from algorithm-centered approaches toward solutions oriented toward practical implementation, supported by sensing, modeling, and human–robot interaction.

Despite the rigor of the adopted approach, this study presents limitations inherent to bibliometric analyses, such as dependence on the selected databases and the employed search strategy. Future research could expand the temporal framework, incorporate new sources, or deepen comparative analyses across specific engineering sectors.

5. Conclusions

This study carried out a systematic bibliometric review of the scientific literature on the integration of artificial intelligence into robotics applied to engineering, based on 1661 publications indexed in Web of Science and Scopus during the 2020–2025 period. Through the combined use of performance indicators and scientific mapping techniques, the temporal evolution, intellectual structure, and main research lines that shape this field have been characterized in detail.

Overall, the obtained results make it possible to understand not only what is being researched, but also how artificial intelligence-driven robotics in the field of engineering is evolving and where it is heading, offering a solid basis for guiding future research and strategic decisions in this area.

The results show an accelerated growth in scientific production from 2022 onwards, which coincides with the consolidation of advanced machine learning techniques and their progressive adoption in engineering robotic applications. This growth is supported by a concentrated editorial ecosystem, where journals such as IEEE Access, Sensors, and Applied Sciences act as main dissemination nodes, reflecting the interdisciplinary and applied nature of the area.

From a geographical perspective, the analysis revealed a clear concentration of scientific production in China and the United States, both in single-country publications and in works with international collaboration. Nevertheless, countries such as the United Kingdom, Italy, Germany, and Canada present high percentages of MCP publications, which highlights their role as key actors in international research networks. These patterns confirm that artificial intelligence-based robotics is a highly globalized field, in which transnational collaboration constitutes a relevant factor for scientific impact.

The conceptual analysis made it possible to identify three well-differentiated thematic clusters. The first, of an applied nature, focuses on operational robotic systems, training, task analysis, and computational modeling, evidencing a clear orientation toward technological transfer. The second cluster, more compact, acts as a connecting element and is associated with sensing and data processing, highlighting the enabling role of these technologies. The third cluster constitutes the methodological core of the field, dominated by machine learning approaches, neural networks, and reinforcement learning, which underpin the development of adaptive and intelligent capabilities in robots.

In a complementary manner, the thematic analysis revealed an evolution from general methodological approaches toward specific engineering applications, as well as the emergence of lines related to human–robot interaction, reliability, safety, and the transfer of models to real environments. This transition indicates a progressive maturation of the field, in which algorithmic advances are increasingly integrated into deployable robotic systems.