Smart Robotics in the EU Legal Framework

3.1 Introduction

This section contextualises the Machinery Regulation (MR) within the EU’s framework for product safety, designed to harmonise health and safety standards for machinery and support trade across the internal market. The MR enforces compliance with Essential Health and Safety Requirements (EHSR) and mandates CE marking and third-party assessments for higher-risk machinery or those incorporating advanced technologies. These assessments are performed by notified bodies to ensure conformity with MR standards.

3.2 The MR and the EU’s ʻNew Approach’ Framework

The MR operates within the EU’s product safety framework, designed to harmonise health and safety standards across the internal market, thereby eliminating trade barriers. This regulation embodies the ʻNew Approach’ to legislation, diverging from the ʻOld Approach’ that incorporated technical specifications directly into legal texts.³² Thus, the MR adopts a technology-neutral stance, setting forth Essential Health and Safety Requirements (EHSR) without mandating specific technical solutions.³³ This leaves manufacturers the freedom to select the most appropriate technical solutions, encouraging innovation and the development of new products. A critical aspect of this approach is the use of technical standards, particularly the numerous Harmonised Standards.³⁴ When manufacturers observe these Harmonised Standards, compliance with the MR’s requirements is presumed, thereby streamlining the process for demonstrating adherence to essential health and safety criteria.³⁵ This method enhances flexibility and fosters innovation by allowing for a broad range of compliant technical solutions.

The MR establishes health and safety requirements for the design and construction of machinery placed on the European market. One of the significant changes in the Machinery Regulation compared to the previous Directive is its explicit intention to cover recent and forthcoming technological advancements in machinery.³⁶ The update to the MD therefore includes provisions on autonomous mobile machinery, connected equipment through the Internet of Things, and some aspects of artificial intelligence where specific AI modules using learning techniques ensure safety functions.

This regulation is crucial for manufacturers and other economic operators like importers, as it ensures that machinery products meet the required health and safety requirements and can be traded without restrictions in the internal market.

3.3 CE Marking and Third-Party Assessments

For end-users of machinery, the most visible sign of compliance with the Machinery Regulation is the CE mark affixed to the product, indicating adherence to the EU’s harmonised framework for product safety. Manufacturers of machinery falling within this regulation’s scope must conduct a conformity assessment process to verify compliance with the essential health and safety requirements. Upon successful completion of this process, the machinery is granted the CE mark, signifying its eligibility for unrestricted trade within the internal market. Depending on the specific type of machinery, a manufacturer may need to comply with multiple regulatory frameworks. For instance, a medical robot could be subject to the MR, the AIA and the Medical Device Regulation (MDR), assuming these laws are applicable.³⁷ Navigating these various rule-sets poses challenges, prompting efforts towards automating compliance analysis.³⁸

A key element of the MR is the role of third-party entities, known as notified bodies, in the conformity assessment process.³⁹ These bodies are crucial, particularly for machinery posing greater risks or embedding advanced technologies like autonomous mobile machinery or AI systems with learning capabilities for safety functions. Notified bodies conduct independent assessments to verify machinery’s compliance with regulatory standards, playing a pivotal role in ensuring the safety and conformity of high-risk machinery categories, including certain robots. The involvement of notified bodies is significant as it also links to the AIA’s applicability criteria, as detailed subsequently.

4.1 Introduction

The MR’s use of terminology, notably the broader classification of ʻmachinery’, encompasses robots, even though the term ʻrobots’ is not explicitly used. While the MR also refrains from using the term ʻsmart robot’, it revises the MD to integrate provisions for autonomous machinery and AI technologies, addressing the evolving landscape of technological advancements in robotics.

4.2 Machines versus Robots

If we are to read the MR as a law on robotics, we first face a definitional challenge, as the ʻR-word’ has no function in the Regulation. Instead, it focuses on ʻmachinery’, meaning an ʻassembly, fitted with or intended to be fitted with a drive system other than directly applied human or animal effort, consisting of linked parts or components, at least one of which moves, and which are joined together for a specific application’⁴⁰. Definitions of ʻrobot’ may not necessarily make use of these exact words, but it is arguably difficult to imagine a physical robot without an assembly of various parts, including a drive system. For example, ISO 8373:2021 describes a robot as ʻa programmed actuated mechanism with a degree of autonomy to perform locomotion, manipulation or positioning’. The definitions of ʻrobot’ and ʻmachinery’ have the drive system (or actuation) in common. However, the definition of ʻrobot’ contains additional requirements. Under the ISO definition, a robot must be programmed, while machinery can also be steered by a human without any programming. Moreover, the definition of robot emphasizss the robot’s degree of autonomy, ie the performance of tasks without human intervention, which is only an implicit possibility for machinery. Thus, robots defined according to ISO 8373:2021 will regularly fulfil the requirements for machinery, but not vice versa.⁴¹ In summary, robots will be subject to the MR, provided that no exclusions apply.

Theoretically, the concept of smart robots includes a range of autonomous machines such as cars, drones, personal service robots, and medical robots. However, in practice, the exclusions listed in Article 2(2) MR encompass a wide array of machinery, including weapons, specific vehicles, seagoing vessels, among others. For instance, robots used in amusement parks, nuclear facilities, or for artistic performances would not fall under the purview of the MR.

For this article, the definitional challenge nevertheless applies. For clarity and brevity, the text discusses the MR as robot-related requirements, even though the regulation also covers a broader spectrum of machinery and related products.

4.3 Smart Robots: AI vs Self-Learning and Autonomy

When regulating smart robots, the MR faces a second terminological challenge similar to that encountered in the broader discourse on AI regulation, especially concerning the AIA.⁴² This challenge is to define AI in legal terms that are technology-neutral, future-proof and practical. Instead of using the term ʻAI’, the MR opts for descriptors such as ʻautonomy’ and ʻself-learning’. These terms are often applied together, particularly in reference to control systems of machinery that exhibit ʻfully or partially self-evolving behaviour or logic designed to operate with varying levels of autonomy.’⁴³ This approach seems aimed at capturing the evolving nature of AI without the limitations of a fixed definition, particularly since the MR predates the AIA.

Autonomy in robotics encompasses the capability of robots to operate and make decisions independently, without human intervention in real time.⁴⁴ This involves various levels of autonomy, from simple pre-programmed tasks to complex, adaptable behaviours that respond to the environment. A common framework for understanding autonomy, especially in the context of autonomous driving, are the SAE levels of autonomy, which range from Level 0 (no automation) to Level 5 (full automation).⁴⁵ While this specific framework is tailored to vehicles, the concept of graded autonomy can apply to robotics broadly, indicating a spectrum from basic automated functions to fully independent operation.

In robotic movements, autonomy can manifest in two primary domains: navigation and manipulation.⁴⁶ Navigation involves the robot’s ability to move through its environment, which could include avoiding obstacles, planning paths, or exploring new areas. Manipulation entails the robot’s interaction with objects, such as picking up items, using tools, or performing tasks that require precise control of robotic arms or manipulators.

One of the key technologies enabling autonomy, especially in learning and adapting to new situations, is machine learning, and more specifically, reinforcement learning (RL). RL allows a robot to learn from interactions with its environment by trying different actions and learning from the outcomes, optimising its behaviour over time.⁴⁷ For example, an RL-powered robot navigating a space could learn to avoid obstacles more efficiently over time based on past encounters or adjust its path based on the dynamic changes in its surroundings. By directly targeting autonomy and self-learning, the MR adopts a relatively concrete set of concepts that should be easy to understand from a technical perspective.

In contrast, the AIA defines an AI system as ʻa machine-based system designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments’.⁴⁸ Both the MR and AIA address the varying levels of autonomy, and self-learning in the MR can be seen as a form of adaptiveness. Influencing the physical environment is inherent in robots and machinery. While the MR does not explicitly refer to ʻinference from input data’, this aspect is likely embedded in self-learning systems, and robots use a variety of sensors for input. Consequently, a robot’s control system that combines self-learning and degrees of autonomy⁴⁹ would likely also fit the AIA’s AI definition. However, whether such a system also falls under the AIA’s definition of high-risk AI is a separate matter, discussed below in Section 7.2.

The practical implications of the terminological distinctions between the MR and the AIA are not entirely clear, potentially leading to additional compliance costs due to the use of slightly different terms for analogous concepts. However, for robot manufacturers and third-party entities, adhering to the MR’s more concrete requirements might prove more straightforward. Nonetheless, ambiguities could arise in cases of minimal autonomy since the ʻvarying levels of autonomy’ are not explicitly defined in either the AIA or MR.

In conclusion, the MR’s strategy of eschewing the broad term ʻAI’ in favour of ʻautonomy’ and ʻself-learning’ enables coverage of a wide spectrum of technologies, including those that allow robots to learn and adapt to their environment. This nuanced approach ensures the MR remains technology-neutral, flexible and able to accommodate future developments in robotics and AI technologies.

The subsequent sections first address the specific requirements that the MR sets forth for smart robots and then explore the interplay between these requirements and those outlined in the AIA.

5.1 Introduction

The MR’s requirements for certain advanced machines are applicable to smart robots and regulate some aspects of HRI. Unlike its predecessor, the MR considers the integration of cobots into human environments. This shift from segregation to integration emphasises the need to rigorously assess and mitigate risks in spaces where humans and robots stay without barriers, or even collaborate or interact. For instance, the force of a cobot’s handshake could unintentionally cause harm, or its unforeseen movement in public spaces might pose a danger to passersby. These scenarios highlight the importance of ensuring robot movements and actions do not compromise human safety,⁵⁰ aligning with Asimov’s first law that a robot should not harm humans.

Considering this, the MR specifically mandates risk assessments and the prevention of hazards stemming from physical contact with machinery, particularly those with moving parts.⁵¹ While this stipulation is expected, it is noteworthy that the required assessment is by no means limited to physical health. It also covers the psychological stress caused by robot encounters.⁵² The MR’s understanding of health and security is therefore a broad one. The design and physical presence of a robot—for example, its height—might intimidate vulnerable individuals like patients, children, or the elderly. This general requirement to assess HRI risks applies regardless of whether the robot is tele-operated or programmed.

The complexities of these assessments become particularly pronounced when considering AI-powered robots, whose capabilities for self-evolving behaviour and autonomous mobility introduce new dimensions to safety and risk management. For smart robots, new provisions are in place, focusing on self-evolving behaviour, and autonomous mobility. The Essential Health and Safety Requirements (EHSR) mandate the identification and management of specific risks for both categories,⁵³ combined with additional requirements detailed in subsequent sections.

5.2 Robot Autonomy

Robot autonomy raises significant safety concerns due to the diminished role of direct human control. Defined in EHSR Section 3.1.1, ʻautonomous mobile machinery’ refers to machinery capable of independently ensuring safety functions, diverging from traditional operations that rely on human operators directly involved in running a machine. This definition aligns with the Parliament’s conception of autonomy, emphasising the machine’s independence from human intervention. However, autonomy in the MR’s context is more narrowly defined, concentrating solely on safety functions.

While the MR accommodates autonomous mobile robots, it sets forth certain basic requirements both for their design and supervision. First, robot engineers must ensure that the technical control system running the robot executes the autonomous function safely.⁵⁴ In addition, a human supervisory function must exist,⁵⁵ serving as a critical yet narrowly focused mechanism for remote monitoring and controlling machinery in autonomous modes. This function allows for essential actions such as stopping, starting, or moving the machinery to a safe state, contingent on the supervisor’s comprehensive understanding of the operational area, whether directly observed or indirectly assessed.

The requirements for robot autonomy and supervision are intentionally broad, necessitating further detail through technical standards. Robot autonomy does not eliminate human involvement, as the MR ensures human participation in both the control system’s design and, to a certain extent, its oversight. However, the MR does not specify the conditions under which human supervision is required nor details the specifics of how such supervision should be implemented. It merely opens the door for supervision, leaving the details of how the supervisory function is staffed and managed to be defined by additional rules (including the AIA), standards, or even user instructions. Therefore, the MR does not provide a comprehensive regulatory framework for autonomous robots; it merely serves as an enabler for such regulation.

This reliance on standards and the mandate for manufacturers to conduct risk assessments before placing a robot on the market or putting it into service⁵⁶ are signs of the MR’s meta-regulatory approach. Meta-regulation refers to a regulatory strategy where the regulation is focused not on dictating specific behaviours or outcomes but on governing the processes through which organisations self-regulate.⁵⁷ This approach leverages the capacity of organisations to monitor their own compliance with regulatory principles, often through internal governance mechanisms, standards or guidelines. Ideally, meta-regulation encourages organisations to develop their own compliance strategies tailored to their operational contexts while adhering to the broader goals or principles set by the regulatory authority. This aligns with the Parliament’s emphasis on the ʻdo not harm’ principle being directed at designers rather than the robots themselves, thereby tasking engineers with evaluating risks and ensuring machinery safety. The MR also accounts for risks that may emerge post-market due to the robot’s evolving and autonomous behaviour, as well as HRI. As mentioned, it does not, however, prescribe specific technical solutions for meeting the requirements.

In the MR, meta-regulation is nevertheless combined with more specific requirements following the logic of ʻcommand and control’. An example is the requirement for autonomous mobile machinery to either operate within a specified zone or possess obstacle detection capabilities.⁵⁸ This is a relatively concrete requirement that can be checked and enforced. However, also this requirement is meta-regulatory, as it does not specify the methods for achieving the requirement. The choice between restricting machinery to a designated zone or equipping it with detection technology is explicitly dependent on the risk assessment. For instance, consider the application of zoning and detection capabilities within a hospital setting, which utilises different types of robots for various tasks. Zoning could be applied to material transport robots, restricting their operation to specific areas like the pharmacy or laboratory wings, where interaction with patients is minimal and staff are trained to interact safely with these robots. This approach ensures that the robots fulfil their designated roles without posing risks in crowded or sensitive areas, adhering to the MR’s safety mandates through spatial limitations.

Conversely, human detection becomes crucial for robots operating in more dynamic environments within the same hospital. Surgical robots or diagnostic robots, although more stationary, might still require the capability to detect human presence to ensure safety during their operation. Similarly, robots designed to navigate public areas of the hospital, such as corridors or waiting rooms, must possess advanced detection capabilities to avoid collisions with patients and visitors. These robots, moving beyond tightly controlled zones, exemplify the necessity of equipping autonomous machinery with the ability to identify and react to humans in real-time, a requirement underscored by the MR to safeguard against the risks associated with human-robot interaction in accessible spaces.

This approach, despite its direct mandates, maintains a meta-regulatory aspect by granting manufacturers the leeway to determine their compliance strategy with safety obligations. By melding these approaches, the MR offers a regulatory framework that is both partly flexible and still stringent.

5.3 AI and Robots with Self-Evolving Behaviour

The MR’s second AI-based category is ʻself-evolving behaviour’ of machines. The notion that systems can have agency is not new,⁵⁹ but self-evolving behaviour requires additional measures. While the MR does not explicitly define the mechanisms of this behaviour evolution, it is reasonable to infer that machine learning or other AI technologies could be instrumental in facilitating the self-evolution of behaviour.⁶⁰ This is in line with Recital 54 of the Regulation which justifies the stricter rules for machinery with self-evolving behaviour because ʻdata dependency, opacity, autonomy and connectivity’ could considerably increase the probability and severity of harm.

The use of machine learning in robots indeed poses safety challenges. Theoretically, a robot’s AI system could enable it to perform new tasks or move in unforeseen ways, potentially leading to harmful outcomes. This ties into the European Parliament’s expectation for HRI to be characterised by directability and predictability.⁶¹ In the context of Asimov’s Laws, particularly the first law (not to harm humans) and the second law (to obey orders from humans) are relevant to the challenges with self-evolving robot behaviour.

A practical concern is that self-learned behaviours might surpass the boundaries of initial risk assessments, pivotal to the safety certification. As a result, stringent design parameters are mandated by the MR to ensure that the robot does not exceed its predefined ʻtask and movement space’.⁶² However, the practical implementation of this requirement raises several questions. Often, the parameters for a robot’s operation might be set, at least partly, by the user, rather than the manufacturer. A robot’s user typically determines the specific tasks the robot should perform and the areas where it can operate. Additionally, a robot’s tasks and movement space are not static; they frequently change based on the user’s commands and intentions. Updates to the robot’s software or alterations to its mechanical configuration, such as adding a new arm or hand, may require the expansion of its operational capabilities. Consequently, the parameters defining allowable tasks and movement spaces need to be adaptable to accommodate such changes. The challenge of addressing these issues falls to future standardisation efforts. While the current provision limiting the robot to its pre-defined task and movement space is somewhat broad, it lays the groundwork for developing detailed mechanisms and protocols to ensure safety compliance.

While ensuring safety makes it crucial to limit what a robot can do, such measures alone may not be sufficient to mitigate all risks. Robots could potentially still unexpectedly execute unintended and potentially harmful actions. Therefore, it is also required to have mechanisms in place to interrupt or correct these actions and retain data about decision-making.⁶³ This provision is similar to Article 14 AIA, as further elaborated in Section 7.

6.1 Robots Responding to People

In robotics, the role of communication extends beyond the mere exchange of information to encompass vital safety functions. The MR stipulates that smart robots should be equipped to ʻrespond to people adequately and appropriately’,⁶⁴ echoing the Parliament’s insistence on predictability within HRI. Communication can take various forms, such as ʻthrough words and non-verbally through gestures, facial expressions or body movement.’⁶⁵ This sets a foundation for more natural and intuitive interactions, but it also implicitly anthropomorphises robots.⁶⁶

This requirement for robot communication is anchored in the MR’s principle of ergonomics. Ergonomics aims to optimise human well-being and overall system performance by designing environments, products, and systems that align with human capabilities and limitations, thereby enhancing safety, efficiency, and even comfort. This broader focus on improving robot functionality for human interaction thus extends beyond mere safety concerns.

6.2 Communication with the Operator

In addition to responding to ʻpeople’, the smart robot should also ʻcommunicate its planned actions (such as what it is going to do and why) to operators in a comprehensible manner’.⁶⁷ The distinction between communicating ʻresponses’ to people generally and conveying planned actions specifically to operators is noteworthy, because the nature and depth of communication varies based on the audience’s role in relation to the robot. While general responses might be necessary for broad situational awareness and safety for all nearby humans, detailed communication of planned actions could be essential for operators responsible for overseeing robot operations. It is also likely that operators can avail themselves of communication channels that are not open for third parties, such as a direct interface with a robot’s control system. Yet, depending on the context it might also be necessary to develop various communicative roles that go beyond ʻpeople’ and ʻoperator’, and which could include patient, family member or other social roles.

Moreover, the MR’s focus on operators in the context of autonomous machinery presents an interesting paradox. The concept of ʻoperators’ traditionally implies direct human control or supervision, yet autonomous machinery, by definition, operates with a degree of independence. This raises questions about how ʻoperation’ is conceptualised in scenarios where machines are designed to make decisions without human input. The regulation seems to anticipate a hybrid model, where autonomous machines perform tasks independently, but where human oversight,⁶⁸ particularly by designated operators, remains a critical safety and operational consideration.

Such considerations underscore the complexity of integrating autonomous robots into human-centric environments. They highlight the need for regulatory frameworks to evolve in tandem with technological advancements, ensuring clear, context-sensitive guidelines for human-robot communication and interaction. The nature of these questions is arguably too detailed and, at least presently, too technical for legislation, so it makes sense to leave these questions to standardisation and allow for innovative solutions, provided that these solutions adequately manage risk.

7.1 Introduction

The interaction between the AIA and the MR is significant, as both regulations deal with aspects of AI, whether integrated into specific machinery or in a broader context. Given that both sets of rules can be applicable simultaneously, it is crucial to examine the criteria that trigger their application and determine if similar issues are handled in comparable manners.

7.2 AIA Applicability Directly or via the MR

The AIA’s applicability hinges upon various criteria that cannot be exhaustively addressed here. However, in brief, AI systems that pose a risk beyond a certain threshold can either be forbidden under Article 5 AIA or regulated as high-risk pursuant to Article 6 AIA.⁶⁹ In the context of robotics, it is essential to differentiate between two types of AI systems within a robot.

First, there could be various AI-based software running within a robot’s system which can be directly regulated by the AIA independently of the MR. In principle, a robot can run any software, such as for migration control purposes or workplace software utilising emotion recognition. Depending on the specific use case, the former example may be regulated as high-risk under the AIA, Annex III (7), while the latter type of emotion recognition is prohibited under Article 5(1)(f) AIA. However, in these cases, the focus is on the software in general, regardless of its installation on a robot or any other hardware (eg a laptop). This means that the conformity assessment of the software (ie the AI system or systems) is conducted separately from the conformity assessment of the hardware (robot or laptop).

This direct application of the AIA needs to be distinguished from the indirect applicability via the MR. Specific machinery components regulated under the MR can trigger the applicability of the AIA due to their criticality. Focusing on the robot as a machine regulated under the MR, certain AI-based systems are categorised as especially critical for health and safety under the MR, triggering the AIA. This is explained subsequently, but an example would be a robot’s emergency stop device using machine learning.

The AIA foresees that an AI system intended as a safety component of machinery (and similar products regulated under the MR) can fall under the high-risk AI category.⁷⁰ This trigger is conditional on the machinery, or its component, requiring a third-party conformity assessment under the MR. If Article 6(1) AIA is triggered, all requirements for high-risk AI systems in Chapter 2 AIA apply. Therefore, it becomes essential to identify MR requirements for third-party conformity assessment.

The MR stipulates a third-party conformity assessment for the safety components or embedded systems of certain smart robots, namely those exhibiting ʻfully or partially self-evolving behaviour using machine learning approaches ensuring safety functions’.⁷¹ As mentioned, this requirement for a third-party assessment triggers obligations for high-risk AI based on Article 6(1) AIA. Consequently, for robots whose safety-related software is considered high-risk AI, the requirements of both the AIA and MR apply to the respective safety component or embedded system.

Therefore, the AIA could apply differently to these—for example, AI-based emergency stop devices—compared to other high-risk AI software running on the robot, such as an AI-based migration software, regulated under the AIA in its own right. The conformity assessment for the AI-based emergency stop system follows the MR, pursuant to Article 43(3) AIA, with additional requirements for high-risk AI systems⁷² applying in addition to the MR’s EHSR.

7.3 Overlapping Human Oversight in MR and AIA

This potential dual applicability leads to the duplication of, at least, requirements regarding human oversight and data recording. Article 14 AIA stipulates that high-risk AI systems must be designed to allow for effective oversight by natural persons throughout their operation. This principle mirrors the MR’s requirement for the supervisability of self-evolving machinery.⁷³ Although the two rule sets differ in detail, both aim to ensure some level of human oversight. Their effect is that, to manage the various autonomous capabilities of a robot, mechanisms for human intervention are in place to correct or halt the robot’s operations should it stray from safe behaviours.

7.4 Overlapping Recording Requirements

Another aspect of partial alignment between the MR and the AIA involves the recording and retention of decision-making data. The MR requires ʻenabled’ recording of safety-related data, with a mandatory retention period of one year,⁷⁴ to ensure machinery compliance with safety standards. In contrast, the AIA mandates that high-risk AI systems automatically record ʻlogs’ of events throughout the system’s lifespan, for very similar purposes.⁷⁵ This highlights a discrepancy in focus: the AIA on ʻevents’ versus the MR on ʻdata on the safety-related decision-making process’. Additionally, under Article 19 AIA, logs must be kept for at least six months or longer, depending on the intended purpose of the high-risk AI system, unless other laws specify otherwise. Thus, while both frameworks require recording—albeit of distinct data types—the retention periods differ. This discrepancy is likely to present practical challenges for companies in ensuring compliance with both regulatory frameworks. In addition, if the retained data are personal, the General Data Protection Regulation (GDPR) limits retention to the period necessary for the purposes for which the personal data are processed, which is a more open-ended assessment.⁷⁶ This limitation is recognized in the AIA’s data retention requirements, but not in the MR’s.

The primary goal of the data-recording mandate is to ensure compliance with established safety standards. However, these data are also potentially relevant for stakeholders determining accountability in the event of robot-related accidents causing harm or property damage. The necessity for thorough investigations in such scenarios is paramount, making these data crucial, for example in the context of liability for damages caused by AI. Access to these recorded data could, however, pose challenges. The MR specifies that data collection serves the sole purpose of proving machinery compliance upon a justified request from a competent authority. Similarly, the AIA mandates data recording to support specific supervisory measures. However, once collected, other legislation might stipulate the disclosure of such data for additional purposes, potentially complicating their intended exclusive use.