1 Introduction

In recent years, machine autonomy has become widely established in mobility contextsFootnote 1, military, and consumer environments (Scharre 2018b). This includes surveillance, logistical tasks, and operations in life-threatening environments or crisis operations (Hudyma 2020; Iliev and Genchev 2020). According to Scharre (2018a), autonomous systems (AS) possess several advantages, such as range, endurance, speed, and coordination, in comparison to existing non-autonomous piloted systems.Footnote 2 However, they also hold certain disadvantages, e.g., in that they lack the flexibility of human intelligence. Since humans and (semi) AS complement each other effectively in many areas, especially in the area where “dirty, dull and dangerous” jobs have to be accomplished (Feickert et al. 2018), combination and coordination are becoming increasingly essential in those contexts (Feickert et al. 2018; Scharre 2018b). An example would be the everyday use of partially and highly automated systemsFootnote 3, which improves rapidly and supports or enhances human capabilities in dangerous or monotonous situations. Currently, the focus lies on procedures with human intervention, meaning that a human remains in control. Nevertheless, it is currently anticipated that more automation and subsequently, advanced semi- or even fully automated systems will exist in the future (Bellet et al. 2019). Improved artificial intelligence (AI) and advanced mechatronic systems will lead to a widespread proliferation of AS (Bagloee et al. 2016). Machine autonomy can only emerge through the interaction of a wide variety of technology fields, which are primarily software-based, but also depend on hardware.

In the future, autonomy will be largely driven by information technology (IT) and the automotive industry (Anderson and Matsumura 2018; Verbruggen 2019). In the military sector in particular, autonomy, or the technology required for autonomy, has various implications. The growing importance of AS is generating intense academic debates on the ethics of AI and autonomy, lethal autonomous weapon systems in general (LAWS), proliferation, and arms control of AS (Charatsis 2017; Ekelhof 2019; Huelss 2019; Morley et al. 2020). Within these debates, especially in the more technology-focusedFootnote 4 ones, AD is particularly prominent for its relevance in society and topicality (Martinho et al. 2021). Not surprisingly, since one of the major trends, especially in the automotive industry, is the vision of driverless cars (Ahangar et al. 2021), which has a long engineering and cultural history (Dickmanns 2002; Robert 2019). Like all AS, autonomous driving (AD)Footnote 5 is a complex technology that is being driven by the industry and numerous scientific fields (Meunier 2019; Zeng et al. 2019) and is clearly an accelerator for research on automation, but also not the only focus. In addition to the automation of the driving process itself, the focus lies on the automated division of the various tasks between driver and vehicle. The integration of the infrastructure, the interactive control and the cooperative, interactive behaviour of the road users are also intensively examined (Bratzel and Thömmes 2018; Flemisch et al. 2015; Martínez-Díaz et al. 2019; Stiller et al. 2018).

Generally, AS need to interact with human beings, which makes the research approach interdisciplinary by nature, including, amongst others, computer science, social sciences, electrical and mechanical engineering. Especially interdisciplinarity is precisely a factor in the diffusion of a technology (Tucker 2012). Within research on automation, engineering-related fields seem to predominate (Bissell et al. 2020; Schaub 2020). An engineering research paradigm prevails, as autonomous systems are hybrids of software, hardware, and mechanical parts (Mallozzi et al. 2019). Measured in terms of global scientific output (published papers), the field of AI and robotics clearly dominates (UNESCO 2021).

In academic discourse, it is widely anticipated that technologies with dual-use potential, meaning that they can be used for both civilian and military purposes, are gaining importance (Shim 2017). Semi-autonomous vehicles are already in military use, witnessed recently in the Gaza Strip (Saballa 2021), indicating that AD can be perceived as dual-use. Relatively to the “trendiness” of the topic of AD, the nexus between dual-use and civilian AS research and development (R&D) has been rarely discussed, particularly with a focus on autonomous vehicles (AVs) (Lin 2016). The same applies to dual-use risks in the field of computer science or software development (Reinhold and Reuter 2019b; Riebe and Reuter 2019), which play a key role in the field of AD (Boulanin and Verbruggen 2017). In comparison to other dual-use goods, such as biological ones, the direct effects of software are often not easily identifiable because the outcomes do not necessarily have to be physically apparent (Reinhold and Reuter 2019a). Generally, software-based technologies are inherently dual-use as they are “incorporated into both peaceful civilian application and military weapon systems” (De Ágreda 2020). Even if research were conducted for a specific purpose or to develop a new technology, the developed technology may be used in opposition to the initially envisaged purpose or objective (Ehni 2008).

Drawing on these thoughts, we conducted 25 interviewsFootnote 6 with developers and researchers from academia and the automotive industry in Germany between July and December 2020 to contribute a German-specific analysis of potential dual-use challenges in the context of AD. In the interviews, we posed the questions whether civilian research on autonomous cars is having an impact on military research and which technical components seem to be of particular interest for military purposes. With this study, we seek to contribute both to the academic discourse and the technical issues evolving from the dual-use problematic determining which technical development, especially in terms of components, can most likely be transferred.

In the following, we present related work (Section 2), incorporating concepts on AD and dual-use. On this basis, we identify research gaps and research questions. In Section 3, our methodological procedure will be illustrated. On this basis, we derive respective findings (Section 4) and discuss the results (Section 5). Lastly, we provide a short conclusion (Section 6).

2 Related Work

In this section, we will present aspects of AD, dual-use, and existing literature regarding dual-use issues in AS, focusing on AD.

2.1 Autonomous Driving

The vision of machine autonomy is relatively simple: machines can make independent decisions that consider their environment. However, to date, fully autonomous machines remain a future vision. The path to this vision entails various technical, social, and ethical questions that remain open. Generally, the research field on AS is fragmented, due to the multifaceted disciplines involved. However, with simultaneous progress taking place within these research fields and various disciplines, this might result in insufficient understanding of the whole AS by the various actors involved. The R&D of AD technologies is similarly being pursued in various technical and social disciplines (Bagloee et al. 2016). Not surprisingly since autonomy itself has “no established definition. It is not a specific technology area with well-defined boundaries, a dedicated academic discipline, or distinct market sector. Autonomy is not even technology per se; rather, it is a property that can be attached to very different types of technology” (Boulanin and Verbruggen 2017). This ecosystem contains various and heterogenous actors such as automotive or robotic companies, universities, or the open source community, which does not necessarily exchange knowledge (Mallozzi et al. 2019). Hereby, machine autonomy advances are primarily driven by AI and robotics (Boulanin and Verbruggen 2017), as subdisciplines of engineering and computer science.

AD is the end point of an ever-advancing automation of mobility.Footnote 7 Through partially and highly automated driving, increasing aspects of driving can be done without human interference. A lot of progress takes place in the field of highly automated and cooperative driving. This is especially important due to the rise in numbers of highly automated vehicles in the foreseeable future. Vehicles have to communicate steadily with each other (V2V) or the infrastructure (V2I). Application cases can be, e.g., platooning, autonomous parking of the vehicle, automatic lane change or intersection management. Likewise, the focus is on the safe and human-centred handover of autonomous driving functions (Bezuidenhout 2013; Faisal et al. 2019; Karg et al. 2021; Malik et al. 2021; Wang et al. 2021) The established vehicle automation levels of zero to five differentiating between various levels of automation. Level zero implies manual operation and at the fifth level, the car itself always decides in every domain, while in level 4, e.g., the driver can still be asked to take over the wheel again. It is important to mention that despite the great promises of car manufacturers in particular, a fully autonomous vehicle (L5) is not to be expected in the near future (Martínez-Díaz et al. 2019).

Regarding connected and automated vehiclesFootnote 8 (CAV), Elliott et al. (2019) note that there are “five areas that lie in the heart of CAV research: inter-CAV communications, security and privacy, intersection navigation control, collision avoidance, and pedestrian detection”. In comparison, Liu et al. (Liu et al. 2018) distinguish three main components of AD technologies: (1) algorithms which include sensing, perception, and decision-making; (2) client systems, incorporating the operating system and hardware platform; and (3) the cloud platform, which involves high definition-maps, deep learning, simulation, and data storage. Despite this partly abstract subdivision, one thing is indisputable: the most important domain is environmental perception, which must be ensured by sensory technology (Campbell et al. 2018). Here, research progress is extremely fast and published results only have a short time span in which they are relevant. The focus lies mostly on the various domains such as sensor technology or algorithms. Overall systemic approaches are underrepresented (Fan et al. 2019).

2.2 Dual-Use and Autonomous Systems

Regarding our research context, we understand dual-use as a term that refers to a technology that can be used for both civilian and military purposes. This can include “materials, hardware, and knowledge that have peaceful applications but could also be exploited for the illicit production” (Tucker 2012) of weaponry or military applications. In general, it should be pointed out that dual-use technologies do not always have to be understood solely as a risk. Generally speaking, it is rather a question of technology assessment. Within the framework of technology assessment, potentials of technical developments may be identified (Schmid et al. 2022). Overall, a framework for assessing potential security risks of nascent dual-use technologies was developed by Tucker (2012), including governability, which includes the parameters of embodiment, maturity, convergence, rate of advance, and international diffusion.

Generally, the military tends to place a high value on performance, survivability, and reliability, neglecting design aspects and cost, whereas civilian development places a high value on low cost, user friendliness, and design (Boulanin and Verbruggen 2017). Principally, it is quite difficult to assess a potentially harmful application, such as sensors or computing units evolving from technological progress. In general, military use does not necessarily have to be a harmful application. Dual-use risks exist in many technologies, ranging from nuclear to life sciences and chemical research, as well as many innovations in IT, such as AI (Harris 2016; Oltmann 2014; Riebe and Reuter 2019). With the rapid improvement of AS, the military has a profound interest to implement civilian R&D into their products or technologies (Hansen and Sauer 2019). According to Kavouras and Charitidis (Kavouras and Charitidis 2020), the flow of information between civilian and military spheres has been reversed, as until the 1960s military research and knowledge mostly influenced civilian research. Further, some emerging dual-use technologies like unmanned aerial vehicles or small drones are increasingly globally accessible for non-state actors (Dahlmann and Dickow 2019).

The dual use-potential of AS might lead to their perception as problematic. As autonomous systems emerge through the collaboration of various technologies (e.g., AI, mechatronics) and different sector fields (e.g., aviation, automotive), they combine different components of knowledge of heterogeneous systems. Therefore, it is relatively easy to transfer technology modules into military solutions (Meunier and Bellais 2019). Although the dual-use nature of computer science is still not the research focus of the discipline (Oltmann 2014), awareness of the issue is growing (Verbruggen 2019), not surprisingly since software in combination with sensors is the main driver and a key aspect for functioning AS (ICRC 2019).

In general, the number of sensors and assistance systems which simplify human-machine interaction (e.g., voice command, smart assistants) or support and sometimes even replace mundane actions (e.g., Advanced Driver Assistance Systems (ADAS)) is increasing. This tendency obscures the borders between machine autonomy, remotely controlled systems, and automated systems (Dahlmann and Dickow 2019). Here, the “civilian market for robotics is considerably larger, many technical components of military applications (e.g., sensors and software) originate from civilian developments” (Dahlmann and Dickow 2019). In addition, due to scale effects of the commercialization in the civil sector, the crucial parts for autonomy (e.g., processing units, batteries or sensors) have improved and become cheaper (Verbruggen 2019). According to Feickert et al. (2018), the progress in the military domain is not at the same pace as in the civilian sphere, which enhances the demand for state of the art solutions. The military has an interest to incorporate civilian R&D, due to potential cost savings, which often outweigh security concerns (Kavouras and Charitidis 2020). This leads to the problem that the technology is hard to control (Sauer 2020). In addition, its general applicability gives autonomous technology high flexibility (Cawthorne and Devos 2020, p. 1573). A study by Martinho et al. (2021) identified 21 ethical issues concerning AV, wherein dual-use ranked 16th, indicating that the subject is not given much attention.

2.3 Research Gaps and Questions

It seems that the field of AD covers a wide spectrum, nevertheless, involved stakeholders often have an incomplete understanding of it, given the small-scale nature of their own research within a complex field. Consequently, we attempted to consult numerous German experts from different disciplines to get a broad overview of possible dual-use technologies in the field of AD. Our empirical study ties in with existing work that addresses dual-use issues and complements the field in terms of AVs, as we have not identified studies in this sector in Germany (gap 1). Furthermore, recent years have shown that developments have been transferred from the military to the civilian sector and that few studies exist on whether this transfer also takes place in the opposite direction, and whether developments from the civilian sector are applied in the military. We question the assessment of R&D that focuses rather on the software side, originally designed for civilian use (Riebe and Reuter 2019). Therefore, our study sheds light on the experts’ assessment of where they see a potential applicability of civil research in the AD sector to the military sector (gap 2). Additionally, we investigate how experts working in the field of AD assess particular technologies. Being interested in which technological components should be further monitored (in Germany), we emphasize the importance of critically observing the transfer of civilian development from AD to the military domain (gap 3).

Inspired by these research gaps, we aim to answer the following research questions (RQ):

  • How do German experts from the field of AD assess the applicability of civil research to the military development sector?

  • Which technology components in the context of AD have a dual-use potential, i.e., the transfer of civil to military applications?

3 Methodology

Since semi-structured interviews with experts offer the possibility of gaining deeper insights into the interviewees’ working processes, we chose an empirical qualitative research design. Instead of relying on statistical representativeness, the qualitative research design focuses on insights of significant correlations derived from a rather small number of cases, thereby allowing a more explicit description of the experts’ concrete perceptions (Bogner et al. 2009).

3.1 Case Selection and Data Collection

Germany currently holds over 50% of the world’s patents for AD (Müller 2020) and is one of the three largest producing countries for passenger cars (Statista 2022), leading us to focus on the country. The currently rather scarce data situation on dual-use and AD can be partly enriched with practice-oriented information through conducted expert interviews (N=25, see Appendix, Table 1; xmax=72 Min., xmin=23 Min.) with mostly German (n=24) developers and scientists in the field of AD. Exploratory field access to the interviewees was facilitated through in-depth internet research (e.g., companies, recent studies, media coverage), personal contacts, and recommendations (snowball sampling) (Moser and Korstjens 2018). Overall, different methodological approaches were chosen to minimize potential bias of non-random samples. However, snowball sampling proved to be an appropriate method to find potential interviewees considering that the field was relatively difficult to access. We selected the interviewees according to their technical expertise concerning our RQ and their involvement in science and industry, thereby covering a broad spectrum of possible individuals involved in the field of AD (Meuser and Nagel 2009). After identifying several potential interviewees, they were contacted via email. 25 of the 72 requested individuals were willing to be interviewed digitally via the data-secure and GDPR-compliant communication software Jitsi Meet. According to the theoretical sampling, it seemed necessary to reach a wide range of respondents with expertise in the studied phenomenon (predefined criteria) without specifying a precise sample size a priori. The sampling size was appropriate once a sufficient amount of information was available to understand the phenomenon under study (data saturation) (Moser and Korstjens 2018). The interviewees were pseudonymized, not financially remunerated, and a consent form was signed by everyone prior to the interview. To facilitate structuring and comparison between interviews, a semi-structured questionnaire, containing 19 open questions (see Appendix, Table 2), was deductively elaborated prior to the interview phase, basing upon existing theoretical approaches (Kallio et al. 2016).

Overall, the respondents work, among others, for German start-ups (n=4), Original Equipment Manufacturer (OEM) (n=2), Tier 1Footnote 9 suppliers (n=5) as well as for research institutions (n=13). They operate in 14 different areas, ranging from AI to environmental perception to cloud architecture or ethics, allowing a diversified insight into various research areas. Although we identified numerous areas, we were not able to cover all. Only one person indicated collaborating with operators in the German military (I9)Footnote 10. Notably, all have a graduate degree (mostly in engineering, including mechanical engineering (n=6), electrical engineering (n=4), or computer science (n=5)), with 13 holding a PhD.

It is worth noting that only two persons identifying as female were interviewed, as it was extremely difficult to find women working in this field. Consequently, a clear gender imbalance in the distribution is apparent, reflecting the reality in the research field. Since the beginning, special attention was paid to finding female interviewees, although this proved to be nearly impossible. Even the snowball sampling process failed to identify additional female interviewees, indicating that the interviewees themselves found it difficult to name female colleagues in the field. After explicitly contacting various institutions and searching the internet for women in practice and research, four were contacted with no response.

3.2 Data Analysis

Once the interviews were conducted, they were manually transcribed and coded by two researchers. The content analysis after Mayring (2000) was identified as an adequate analysis method, due to the possibility of category formation. The abductive approach allows to derive theoretical tendencies from empirically collected data and to incorporate available literature in the field. Key findings – relevant for our RQ – were categorized into categories (see Appendix, Table 3). The categories were defined by the researchers and a total of 1971 codes were derived from the interviews using the software program MAXQDA. Taking scientific quality criteria into account, the interviews were conducted and evaluated jointly by two authors, both skilled in computer science and political science, obtaining a multi-perspectivity (Steinke 2004). The interdisciplinary approach of consensus coding was helpful to assess qualitative statements more objectively. As the coding scheme illustrates, the interviews provided a large number of valuable insights. Given the limited scope of this paper, not all findings can be discussed here.

4 Empirical Findings

Interested in answering the two RQ, we identify which technologies from the field of AD are perceived as potentially applicable from the civilian to the military sector by the interviewed experts. Further, our analysis indicates what technologies possess a dual-use potential in AD.

4.1 Potential Applicability of Civilian Research in AD on Military Domains

Throughout the interviews, we postulated that civilian technology might potentially be used for military applications, asking the interviewees whether one’s research might be used for ulterior purposes, such as the military, and how this influences their everyday work. In general, the interviews indicated that many people involved in the field are often unaware of potential transfer between civilian research and the military, including existing dual-use potential. The technologies on which they are currently working on have scarcely been identified as a concrete risk (I12). I7 (71) mentioned that they currently do not see a “hardcore linkage between autonomous vehicles and possible benefits of the military”. One explanation for this might be the high level of abstraction, as it is challenging for many to picture how their own small-scale research may be adopted in military applications (I17). Even a personal embedding in the wide-ranging R&D process of autonomous vehicles seems excessively complex for many, as multiple different fields of research (including deep learning, cyber security, mapping, ergonomics, sensors, and Human Machine Interface (HMI)) exist (I1). Consequently, it appears difficult to situate oneself in the overall field of AD (I19) and to reflect on the issue independently, without someone suggesting it. Another possible reason for not linking one’s own research to military applications, could be the difficulty of admitting that one’s own research can possibly be used within the military. Even in some cases, cooperation with the military is prohibited by the employer (I2). In total, only one person stated that they collaborated directly with stakeholders from the German military (I9).

In total, nine interviewees considered the possibilities that a transfer between certain civilian findings and the military either had already taken place or would certainly be possible in the near future (n=16). Regarding the time needed to transfer certain knowledge, estimates varied between several months (I12), to a few years (I14), to ten years (I18), whereas most people tended to predict a rather short period of time. These different assumptions regarding time illustrate that, so far, no cooperation is known to the interviewees. Although it seems difficult for some to imagine a transferability of their own research, the majority of respondents (n=24) identified at least one technology from AD where they could envision a concrete transfer from the civilian sector to the military. With a temporal component in mind, some interviewees indicated that they believe that a transfer between the military and civilian sector is already taking place (n=9) and that the military is significantly advanced in its own developments (n=12), and therefore does not specifically rely on civilian research (I1). One reason for the technical lead could be the availability of large financial resources. Whether this is truly the case is difficult to verify, as access to information on the military is scarce (I23). It becomes apparent that the interviewees do not know whether there are (official) documents reporting on already existing collaborations between the civil sector and the military sphere, where they could potentially obtain information. The different perceptions and assessments illustrate that a varying dual-use awareness can be identified among respondents.

Regarding their own research, the assumptions differ slightly. In contrast to those not perceiving direct transferability of their own research, others (n=10) consider it less abstract and identify an easy transferability of their own research: “The technologies we produce are basically for road vehicles – military vehicles also use roads. The technologies can be developed further, even for off-road terrain” (I6). This illustrates that irrespective of the work area, some respondents potentially consider a transferability of their own research and others do not.

In summary, regarding the question of whether an exchange between AD technologies and the military is already taking place, most respondents indicated that a transferability of some technologies is easily possible, as certain similarities between autonomous civilian vehicles and military tanks can be perceived (I10). Although most similarities lie between AD and tanks, developments may be applied in other military domains, such as drones (I16). Basically, many respondents indicated that the extent of knowledge exchange between the civilian and military sectors remains uncertain, since little insight into concrete collaborations exists (I17), making it difficult to identify military technical requirements and evaluate on a concrete dual-use potential in the field. Further, interviewee 6 emphasized that even if certain developments could be used mutually, complex process flows make exchange difficult and thus reduce the dual-use potential.

4.2 Transfer of Technical Components from Civilian AD Sector to Military Sector

While we first asked about applicability between the civilian AD sector and the military domain in general, section 4.2 now focuses on concrete technical components that can potentially be applied. In response to the question of which technologies from civilian research in the field of AD could be applicable to the military, the interviewees replied fairly similarly, identifying sensors and environmental perception as the technologies most easily transferable from the civil to military sector. The following figure (see Fig. 1) visualizes technologies mentioned by the interviewees, which they consider to be potentially transferable from the civilian to the military domain. The respondents mentioned sensors (n=12) and environmental perception (n=11) the most, followed by algorithms (n=5), AI (n=3), and actuator technology (n=3). Interviewee 2 explained that environmental perception is crucial in order to determine object location in a precise way. In recent years, a steady improvement in laser and radar sensor technology and cameras has been seen, providing high resolution images even at long distances (I14). In military operations, accurate object recognition over long distances is necessary to ensure, for example, that a correct target can be detected and identified (I24). Interviewee 11 considered environmental perception to be vital, to ensure AS to move freely and identify targets. However, the actual attack seems to remain human operated. Interviewee 19 stated that militarily strong countries like Israel already seem to have made great progress in the development of thermal cameras and sensor technology. The interviewees stated that well-developed environmental perception systems already appear to be used for various purposes, including the identification of individuals in desert regions (e.g., in Israel) (I19) and surveillance (e.g., in China) (I2). Since such new developments intervene in various spheres, such as the individual’s private sphere, a critical perspective on them should be maintained in the future (I2).

Fig. 1
figure 1

Technologies with most potential to be transferred to the military sector (according to interviewees). (Source: own research)

In certain contexts, it seems relevant that vehicles navigate autonomously, “requiring them to observe their environment, understand and reduce what other actors are doing, so that they can respond and cooperate” quickly (I11). Interviewee 1’s assertion is similar, considering perception as essential for the development of AS, meaning that the entire sensor technology thus plays an important role for the military. Interviewee 12 considered numerous areas, in which sensors could be applied in: “Sensors can be used in a variety of ways. The sensors do not care what they are used for – no matter whether they are used on land, at sea or in the air. Even space missions are possible with radar and optical sensors.” Particularly radar and LiDAR (Light Detection and Ranging) serve for object detection, distance control, and collision avoidance (I24). The farther away signals can be received from the LiDAR, and the better the night vision cameras are, the easier it is for the military to confront potential adversaries (I1). On the one hand, potential areas of application were specifically mentioned; on the other hand, some respondents expressed the need for adaptations in the transfer of sensors and environmental perception technologies to the military. They agreed that quick and easy adaptation is possible through interfacing with AI and algorithms (I12). Regarding AI, a different input needs to be defined, meaning how the AI is trained. Ostensibly, the difference between an AI learning to recognize flowers, for example, or faces seems negligible (I25). Deep learning techniques from AD thus seem to be particularly interesting and helpful for the military (I20), although they have probably already conducted intensive research on machine learning itself (I6). Interviewee 15 expressed seeing a one-to-one transferability of sensors, AI, and other methods from the civilian to the military sphere (I15). This case might illustrate the inconsistency among the interviewees regarding the perceived potential for transferability. Certainly, this is partly related to the fact that the military is a large sector (I20), mostly conducting confidential research with insights into the military’s development processes being scarcely accessible, thus making it difficult to identify a realistic risk (I24). Interviewee 17 acknowledged that “if I had a little bit more knowledge of the military, I would probably think of other interfaces”.

Apart from environmental perception, image processing and the transfer to other systems are important to process data (I19). The way in which two or more systems communicate with each other is also relevant for the military. According to I15 (61), technical approaches from car-to-car communication have the potential to be applied in military applications as well. Another area mentioned, which relates to environmental perception, concerns perceptual algorithms capable of estimating where objects will be located in the future. For weapon systems, accurate prediction is crucial to precisely estimate the location of a targeted object within a few seconds (I14). In addition to the transfer of civilian development to the military sphere, other developments have been initiated by the military (I25). GPS, for example, which had originally been developed by the U.S. Department of Defense, now serves for navigating and determining one’s own position. If GPS fails, e.g., in extraordinary situations such as conflicts or natural disasters, systems have to operate on different means of navigation such as tracking nearby objects like trees. Since GPS is one important element for AD, alternative possibilities for high-precision localization are being researched. A possible transfer would be conceivable in this case as well (I3). Precise localization is only of use, if the control processes in the car or system are precisely defined. Control engineering seems to be a complex task, especially in difficult terrain, which is usually the case in conflict regions. Despite challenging outdoor conditions, systems have to be able to target accurately, further indicating their relevance for the military sector. Interviewee 24 mentioned that they believe “control technology is underestimated” and that “not many people have this aspect in mind”.

With regard to rather mechanical and hydraulic processes that are transferable from the civilian to the military sector, actuator technology was mentioned a few times (n=4). Actuator technology plays a major role in automation engineering, helping to understand, for example, how a tank’s hydraulics work in different situations (I18). Actuator technology in most cases appears to be relatively advanced compared to sensors, making it more easily transposable, albeit further research is also required (I12).

In addition to rather small-scale developments, the entire autonomous car can be equipped with a weapon and used in warlike situations, for example (I17). When it comes to route planning then, approaches from civilian trip planning, using high-definition map material (I1), can be adopted, although it tends to be more difficult in military operational terrain since roads are often damaged or non-existent (I3). Other concerns were raised about requiring extensive infrastructure and detailed modelling of the environment in order to provide accurate route planning: “I cannot imagine that in military scenarios such stabilities will be available in the foreseeable future” (I12). Looking at training data, a striking difference between civilian and military requirements appears to persist (I18). Then again, sparsely populated areas with few hindrances, such as pedestrians, appear to be more feasible for AS: “If I want to let a missile, a drone or a tank run in an automated way, I probably have less around me, therefore being able to better control the scenario. One does not have to manoeuvre it through any pedestrians or anything similar – and if so, the complexity of the situation seems easier to grasp, from my perspective” (I4).

In general, it becomes apparent that some development processes, in both hard- and software, are very generic and not tied to AD. Especially in software development, packages are produced rather universally (I18). Therefore, they are easily transferable to other systems (I19) and do not differ much between companies and sectors (I18). When looking at possible transfer, another noteworthy aspect is that military equipment such as tanks can carry significantly more weight than smaller autonomous vehicles in the civilian sector. This means more sensors and technology can be loaded (I15).

5 Discussion

5.1 Dual-Use Technologies in AD

Focusing on the possible applicability between civilian R&D on AD and the military sphere, we were first interested in the extent to which this is assessed by the consulted experts focusing on Germany (RQ1). We found out that numerous technologies that make AD possible have the potential to be used in non-automotive contexts. This does not necessarily imply that its usage must be a genuinely military one. In retrospect, technologies could always be used for multiple purposes. But the progress to extend the capabilities of AS in combination with IT-related progress is a rather unique situation. Of course, it is not the case that certain parts which are related to unmanned aerial vehicles can specifically and immediately be used in an autonomous system on the ground – but the building blocks that are used to solve domain-specific problems can be adapted. Here, environmental perception in particular stands out, as it represents a combination of software and hardware on the one hand, and a general application in AS with broad applicability on the other.

Furthermore, a clear observation is important to scale effects in autonomous technologies. Especially R&D budgets in the civilian sphere will lead to cheaper sensors and off-the-shelf solutions, which consequently will lower the cost for more sophisticated military applications or solutions. Future research should focus on this as well as on different types of machine autonomy. Of course, there will always be different requirements for the various areas of operation (air, ground, maritime, space), but it can be assumed that the overall systemic solution competence is at least transferable. The already thin line between the military and civilian domain seems to be blurring further. Therefore, the question arises whether the separation into different types of AS is still applicable when it comes to dual-use and AS. If we apply our findings into the decision framework developed by Tucker (2012, p. 74) regarding governability, which includes the parameters of embodiment (hybrid), maturity (moderate), convergence (high), rate of advance (high) and international diffusion (high), we conclude that AD can be classified between low and medium governability. As Tucker (2012) has pointed out, soft-law and informal measures, such as international standards and codes of conduct, should be the focus of attention here, given the low to medium governability identified.

Secondly, we posed the question (RQ2) which technology components possess dual-use potential in the context of AD. We have observed that the growing importance of software in cars seems to intensify dual-use in this field overall. As shown by the interviewees, the different technological solutions for environmental recognition, or the sensors being used, are a good example for this observation (Campbell et al. 2018). It can also be assumed that AVs will not operate in a separate manner from AS but will be embedded within a system of systems. This means that in addition to the development of AVs, the overall systemic software embedding will also become an important parameter (Schwartz and Reuter 2020).

Similarly, the important factor of standardization should not be underestimated (Sandl 2021). In the interviews it became apparent that this is of utmost importance. Standardization is currently being developed, and without it, autonomous vehicles will rather be difficult to integrate in everyday life. Of course, the military will not simply adopt civilian standards and solutions, but it must not be neglected that the military solutions must be able to interact with the civilian system, especially in logistics (software- and traffic-wise).

5.2 Recommendations

This paper concludes that respondents mostly perceive that technologies from the civilian sector of AD may be transferred to the military sector, potentially causing certain challenges. So far, it seems that issues in the development of AVs are currently not in the focus of research. Therefore, we recommend further research to be conducted in this area, especially from an engineering point of view. Additionally, awareness raising measures shall be initiated in the different development processes and dual-use aspects and the ethical implications arising from the development of building blocks for AS should become a much more prominent aspect. To critically question or accompany such developments, a far more interdisciplinary approach can be helpful. Hereby, a broad range of social science and engineering disciplines, e.g., AV developers, engineers, regulators, and ethicists should be involved, aiming for an open discussion (Rhim et al. 2021). In general, we noticed that the understanding of dual-use varies greatly and that widely differing terms are used with regard to dual-use. Due to multiple research disciplines from academia and the private sector, a common understanding of the terminology “dual-use” should be elaborated to have a better understanding of key terms used in the debate. A possible and rather practical starting point for addressing this issue could be expert panels on existing conferences, interdisciplinary workshops, and industry fares. There, a discussion on whether a transfer of know-how and technologies should be permitted and, if so, how it can be arranged in an ethically justifiable and transparent manner, should take place.

In summary, this study represents a contribution to a sub-field of dual-use discussions, focusing on AD. Now that potential dual-use technologies in the field of AD have been identified, a concept for technology impact assessment and potential monitoring by the research community and other stakeholders involved may be devised to better observe which technologies can eventually be used in the military domain. Here, a framework worked out with stakeholders involved in the field of AD would simplify the progress. When technological aspects of concern are highlighted, actions on and awareness for dual-use issues in general are more feasible to implement.

5.3 Limitations and Future Work

Based on the conducted interviews with Germans experts, we offer an excerpt of reality in a non-representative manner. We are aware that our study does not generalize overall technological developments in the field of AD, but rather represents the dual-use risk assessment of 25 experts in the field. Firstly, because our participants’ demographics are skewed white, well-educated employees holding a good professional position and secondly, because Germany differentiates from other countries in terms of its military-industrial complex. Certainly, other technical fields and countries show a different transferability between civilian developments and the military. Further studies may complement our sample by looking at different countries and sectors in order to achieve comparability.

Further, based on our educational and socioeconomic background, we are aware that we hold personal beliefs in certain sensitive topics and are not completely objective towards military developments. We tried to avoid any bias during the interviews and to continuously question our own limitations. In general, we hope that our findings lead to further discussions on how technological innovations from the civil sector may be transferred to the military, and what effects this may have. Here, studies on technological impact assessment of the identified technologies would be of interest. It remains to be mentioned, once again, that we personally consider the scientific monitoring of possible dual-use technologies in the field of AD to be important. However, we are aware that this is certainly not universally valid and thus debatable. It should also be emphasised here that the background of both authors is technical peace and conflict research, in which dual-use processes are observed very critically by nature. Potential opportunities that are equally conceivable here may therefore be underrepresented. While this paper focuses on potential dual-use technologies, aspects regarding dual-use awareness will be addressed in a separate paper using the same dataset (Schwartz et al. 2022).

6 Conclusion

The study illustrates that German experts in the field of AD consider many developments in the field of AD to be potentially applicable to the military sector. Although the transferability of certain technologies appears to exist, some people may be reluctant to acknowledge their own research possibly being transferred to the military sphere, perhaps in light of the frequently rather negative perception of the military. Additionally, most interviewees do not seem fully aware of how their own development processes can be applied in the military sector. This indicates potentially underestimated dual-use awareness among respondents involved in the field of AD. Further, the small-scale nature of the research and the lack of embedding it in the overall research field seem to be partly responsible for the missing or little awareness. Since interviewees have different professional backgrounds, very diverse aspects were mentioned for a potential transfer, which is why an overview of all areas of AD appears elusive. Nevertheless, sensors and thereby environmental perception have been identified as the two fields with the greatest potential for transferability. When we assume a transfer to the military sector to require critical examination, then a need for increased monitoring of these developments exists, in order to understand how they may impact the military sector and how developments may potentially be applied in conflict situations. The presented study showed both civilian and military developments influencing each other’s development processes in various ways, and that precise insights into possible cooperation are difficult, as much remains classified.