Beyond the general HRI principles described above, the DAC-HRC architecture incorporates two core principles coming from our current understanding of the origins of human collaboration: interdependence and joint intentionality.

Joint intentionality refers to the shared mental states and cooperative activities that arise when individuals engage in collaborative endeavors. Research on social cognition posits that shared intentionality is a unique feature of human cognition, setting us apart from other primates Tomasello et al. (2005). It manifests in the form of shared goals, joint attention, and mutual knowledge among individuals working together. For instance, when two people collaborate to lift a heavy object, they share a common goal (i.e., moving the object) and are aware of each other’s intentions, roles, and actions.

This concept is particularly relevant in the context of Human-Robot Collaboration (HRC), as it emphasizes the importance of mutual understanding, communication, and coordination between human workers and their robotic partners. In industrial HRC, developing systems capable of exhibiting joint intentionality is essential for ensuring more efficient and safer interactions between human workers and robots. In this context, developing shared intentionality in artificial and hybrid collaborative systems would imply the ability to (1) detect and predict human intentions, actions, and goals, (2) communicate its intentions, actions, and goals to human workers and (3) coordinate and adapt its behavior based on the shared goals and the feedback from the ongoing collaboration.

Interdependence is another foundational aspect of current theories of the evolution of human cooperation Tomasello et al. (2012). It refers to the reliance of individuals on one another to achieve shared goals or complete tasks, and its key role in vital tasks for early humans such as obligate collaborative foraging Tomasello (2009), O’Madagain and Tomasello (2022). Applied to the context of HRC, interdependence implies that both human workers and robots depend on each other’s actions, skills, and knowledge to execute tasks effectively. The Interdependence Hypothesis Tomasello et al. (2012) suggests that interdependence fosters cooperation, as it encourages individuals to align their goals, share information, and coordinate their actions.

In HRC, task interdependence between humans and robots can motivate the design of systems that (1) recognize the skills and capabilities of human workers and adapt their behavior accordingly, (2) share task-related information with human workers, facilitating mutual understanding and efficient task execution (3) respond to changes in the task or environment, adjusting their actions to maintain effective collaboration.

To illustrate, consider a collaborative robot in an automotive assembly line. The cobot could be designed to recognize the specific skills of the human worker, such as their proficiency in installing certain parts. Based on this recognition, the cobot could adapt its behavior to complement the worker’s skills, perhaps by preparing the necessary parts or tools for the worker’s next task. Moreover, the cobot could share task-related information with the worker, such as the sequence of assembly steps or the status of the parts supply, facilitating mutual understanding and efficient task execution. Finally, the cobot should be able to respond to changes in the task or environment. For instance, if a part is missing or defective, the cobot could adjust its actions, perhaps by fetching a replacement part or alerting the worker to the issue.

In understanding and implementing these core principles of joint intentionality and interdependence, it becomes apparent that a sophisticated cognitive architecture is required: one that not only comprehends human social behaviors but also adapts and responds to the dynamic nuances of industrial settings. This necessity brings us to the Distributed Adaptive Control (DAC) approach, which provides a validated and biologically-inspired framework. The DAC approach, with its layered control system and emphasis on adaptability, is ideally suited to embed these principles into the fabric of human-robot collaboration. As we transition to exploring the DAC-HRC architecture, we will see how each of its specialized modules is designed to operationalize the principles of joint intentionality and interdependence, thereby creating a harmonious and effective collaborative environment between humans and robots in industrial settings.

Task planners play a pivotal role in robotics, especially in enabling robots to adeptly navigate complex and unpredictable environments. In domains like electronic waste recycling operations, the ability of robots to perform a range of tasks, from sorting to processing diverse types of devices and components, hinges on sophisticated task planning mechanisms Alami et al. (2005). The cornerstone of contemporary task planning in robotics is the use of finite-state machines (FSMs), revered for their simplicity and intuitiveness in modeling robot behavior amidst uncertainty Foukarakis et al. (2014).

Finite-state machines are essentially mathematical constructs encompassing a finite set of states, transitions between these states, and corresponding input/output events. This structure empowers robots with the ability to efficiently adapt their behavior in response to varying conditions, a feature crucial in the fluctuating environment of a recycling plant. The inherent simplicity of FSMs, however, can be a limitation when dealing with more complex behaviors.

To address this complexity, hierarchical finite-state machines (HFSMs) have emerged as a potent solution to orchestrate complex robot behaviors. HFSMs represent behaviors in a layered structure of FSMs, where each level corresponds to a specific subtask or behavior component Johannsmeier and Haddadin (2016). This hierarchical arrangement facilitates a modular and scalable approach to task planTask Planner, the Socially Adaptive Safety Engine, the Worker Model and the Interaction Managerning. By breaking down overall robot behavior into manageable subtasks, HFSMs offer a tailored solution to the multifaceted tasks encountered in electronic waste recycling. This approach not only enhances the robot’s efficiency and adaptability but also allows for easier integration and updates to the task planning system as recycling requirements evolve.

Moreover, the incorporation of human-in-the-loop methodologies in task planning signifies a significant evolution in robotic systems. This approach involves integrating human feedback and inputs directly into the robot’s control mechanism, enabling a more dynamic and adaptable interaction between humans and robots Raessa et al. (2020). In the context of electronic waste recycling, this means that robots can be more responsive to human operator’s preferences and needs, thereby enhancing collaboration efficiency and safety.

In implementing HFSMs, the Task Planner module within the DAC-HRC cognitive architecture embodies these principles, leveraging the hierarchical structure to manage complex tasks while remaining adaptable to the diverse challenges presented in electronic waste recycling. The module’s design allows for seamless incorporation of human inputs, ensuring that the robotic system is not only responsive but also attuned to the needs and preferences of different human workers. This integration of advanced HFSMs within the DAC-HRC architecture illustrates a commitment to developing robotic systems that are both technically proficient and collaboratively effective in complex industrial settings.

The DAC-HRC architecture’s Interaction Manager advances the paradigm of multi-modal non-verbal communication, pivotal for intuitive and effective human-robot collaboration in industrial settings. In the human-centered HRI paradigm, an essential aspect of implementing a successful and effective HRI is building a natural and intuitive interaction Wang et al. (2022). In recognition of the importance of non-verbal communication modalities, particularly in noisy industrial settings, the Interaction Manager eschews auditory channels in favor of gesture-based and tablet-based interfaces.

Gestures serve as a fundamental form of human communication, making them ideal for conveying rapid commands in human-robot interaction (HRI) Vouloutsi et al. (2020), Pezzulo et al. (2019). Gesture-based communication harnesses the natural propensity for humans to use physical gestures, thereby facilitating a more immediate and universal form of interaction Liu and Wang (2018), Wang et al. (2022), Peral et al. (2022). The Interaction Manager incorporates a repertoire of shape-constrained gestures Alonso-Mora et al. (2015) tailored to the communication needs specific to the HR-Recycler’s project, which facilitates natural and intuitive interactions without extensive training Vouloutsi et al. (2020) while also ensuring accurate recognition and interpretation by the robotic agents Peral et al. (2022).

To complement gesture-based interactions and cater to scenarios necessitating more detailed information exchange, the architecture also integrates tablet-based communication. This method leverages interactive applications, which, while commonly used for teleoperation Yepes et al. (2013), Best and Moghadam (2014), Luz et al. (2019), are repurposed in the DAC-HRC to enhance the human-robot bond and situational awareness Goodrich and Olsen (2003). The tablet application provides a direct interface for the human worker to receive updates on the robot’s internal state and environmental interpretations, aligning with key HRI principles Goodrich and Olsen (2003), Adams (2005).

The combined use of gesture and tablet-based interactions by the Interaction Manager represents a state-of-the-art approach to non-verbal HRI. It successfully navigates the challenges of noisy industrial settings, where traditional verbal communication is untenable, and establishes a robust, adaptive, and user-friendly communication system conducive to the dynamic requirements of the HR-Recycler’s operations.

A robot must not endanger a human under any circumstances. This premise, already formulated by Isaac Asimov in his famous “Three Laws of Robotics,” is crucial for any robotic installation but especially for those promoting interaction and collaboration between humans and synthetic agents Asimov, (2004). Importantly, in industrial settings interactions must be designed bearing in mind that robots occasionally will represent a source of danger to humans because of the tools they employ to perform hazardous tasks such as cutting, hammering, or moving heavy objects. Safety measures need to be implemented both according to the work to be performed and the human demands Van Wynsberghe (2020).

In the context of the WEEE recycling factory, humans and robots have to interact with a variety of different objects and tools and realize many changing sequences of actions in order to successfully complete their tasks Axenopoulos et al. (2019). Although each of the robotic components has its own built-in safety mechanisms and corresponding certified ISO safety measures, the interaction of all these elements together will require an additional layer of control that can adapt their behavior to the requirements of the hybrid recycling plant while following human-centered design principles. This layer of control is the Socially Adaptive Safety Engine (SASE).

The Socially Adaptive Safety Engine within the DAC-HRC architecture goes beyond mere harm avoidance to actively promote cooperation Freire et al. (2020a). The SASE not only adheres to basic safety principles but also engages in more flexible adaptation of the robot’s behavior to the preferences of the human worker, thus fostering a more cooperative and harmonious human-robot interaction. It also reflects the principles of shared intentionality and interdependence, as it involves a mutual understanding between the human and robot about the safety measures and reliance on each other’s capabilities to maintain safety during the disassembly process.

The goal of the SASE is to promote human-robot cooperation by building safer, more trustworthy, and personalized interactions with human users Kok and Soh (2020), Christoforakos et al. (2021). It does so by regulating and adapting the robot’s behavior to the particular human preferences of every user Senft et al. (2019), Edmonds et al. (2019). In this way, it also serves as an extra layer of security for the system by integrating contextual information from the environment and using it to prevent potentially harmful situations Yang et al. (2018). At the heart of this approach is the implementation of an allostatic control system Sanchez-Fibla et al. (2010), Guerrero Rosado et al., 2022. This system aims to ensure harm avoidance and promote cooperative behaviors, which are two fundamental aspects of ethical machine behavior Freire et al. (2020a).

In essence, the Socially Adaptive Safety Engine encapsulates the ethos of the DAC-HRC architecture—prioritizing human safety and introducing dynamic adaptability, thereby exemplifying a model of responsible and responsive artificial intelligence in industrial settings.

User modeling systems rely on data gathering to create user models, either explicitly or implicitly Luna-Reyes and Andersen (2003). The integration of novel machine learning techniques has significantly enhanced the capabilities of these systems, steering them towards more data-driven strategies Kontogianni et al. (2018). One emerging technique being implemented in these data-driven user modeling practices is the Digital Twin concept, which generates or collects digital data representing a physical entity, emphasizing the connection between the physical and virtual counterpart through real-time information flow Bruynseels et al. (2018), Negri et al. (2017).

Digital Twin technologies have been applied in various contexts, such as healthcare Croatti et al. (2020) and human-robot interaction Wilhelm et al. (2021), Malik and Brem (2021), Wang et al. (2020). In industrial settings, Digital Twins have been utilized for tasks like interactive welding, bridging human users and robots through bidirectional information flow, and benefiting novice welder training Wang et al. (2020), Jokinen and Wilcock (2015). However, these data-driven approaches raise concerns regarding big data management, privacy, and trustworthiness, especially when applied to sensitive fields Kumar et al. (2020). The Human-Centered AI paradigm aims to address these concerns by prioritizing methodologies that meet user needs while operating transparently, delivering equitable outcomes, and respecting privacy Xu (2019), Riedl (2019). This approach also aligns with legislation such as the European General Data Protection Regulation (GDPR) Kloza et al. (2019).

The Worker Model module of the DAC-HRC cognitive architecture follows such human-centered design principles by maximizing functionality while minimizing the amount of data gathered from the user Xu (2019), Riedl (2019). This design strategy ensures that the Worker Model respects user privacy while still providing effective support for human-robot collaboration in the disassembly of WEEE devices. The main goal of the Worker Model is to collect, process, and store all the relevant information regarding each user of the system and integrate it into one single, coherent data structure. This information is used by the DAC-HRC architecture to flexibly adapt the human-robot collaboration paradigm to the human partner. In other words, the Worker Model creates a virtualization of the human worker that allows the collaborative architecture to dynamically adjust its parameters to ensure a personalized interaction.

In essence, the Worker Model’s integration into the DAC-HRC architecture not only enhances the adaptability of the human-robot collaboration paradigm but also embodies a human-centric focus into the design of these new technologies Xu (2019), Riedl (2019).

The rest of the chapter is organized as follows: In the following section section, we first describe the aCell, a specific experimental setup designed for the collaborative disassembly of WEEE devices. We then continue describing in detail each of the components of the DAC-HRC architecture along with its interactions. We proceed with a report of the main results showcasing the functionalities of the architecture across the different tested use cases, and conclude with a discussion of the main outcomes of the study, its limitations, implications, and future work.

The DAC-HRC’s Task Planner module is conceived as a human-in-the-loop hierarchical finite state machine that encompasses all disassembly steps of all devices, as well as the error-handling protocols. The Task Planner (TP) has been developed to ensure robust orchestration of various components contributing to the disassembly of WEEE devices within the aCell system. The objectives of the TP are to coordinate the proper disassembly steps for each device, organize the disassembly procedure and robot-worker interleaving, centralize task-related information among the DAC-HRC modules, and implement safe and robust error-handling protocols. The TP reflects the principles of shared intentionality and interdependence, as it involves a mutual understanding between the human and robot about the sequence of tasks and reliance on each other’s capabilities to complete these tasks. The TP integrates and coordinates low-level sensorimotor information (coming from the computer vision and robotic components of the aCell) with high-level information about the task and the interaction (coming from the upper control layers of the architecture). Therefore, within the TP’s HFSM, we find states with different levels of abstraction and description. The Task Planner operates at five levels:

• Task Planner. This level corresponds to the highest level of abstraction, which contains the state machines (SM) of all 4 Devices. It also contains the functionalities that deal with continuous status reports, as well as direct human interactions (through the Interaction Manager, or IM) that allow the TP to be suddenly interrupted by the worker.

The Task Planner is implemented with the Smach-ROS python library, which allows seamless integration of HFSMs with ROS-based communication protocols Bohren and Cousins (2010), Pradalier, (2017). Crucially, internal data structures allow the conveying of information received from vision (response of a ROS service) to the robot (goal of a ROS action). In an SM, the transitions between states depend on the outcome of each State after having been executed. An example of a Task can be seen in Figure 3. In general, an outcome “succeeded” will make the SM transition to the next Task or Action (depending on the level). The outcome “aborted” will engage the error-handling loop (see section Error Handling below), which asks the worker for feedback, and transitions to different states according to the worker’s decision (e.g., the robot tries again, or the worker finishes the Task and then the TP moves to the next Task). The hierarchical structure of the FMS can be achieved because all SMs are treated as States too, inheriting their properties.

The Interaction Manager module plays a vital role in facilitating efficient communication and interaction between humans and robots. To achieve this, the module integrates multimodal channels of communication, ranging from audiovisual interfaces to embodied non-verbal communication. To account for high levels of noise and equipment worn by workers, verbal communication was excluded from the communication repertoire. The two main modes of interaction, gesture-based communication, and tablet-based communication, have been chosen to address the noisy industrial environment and safety concerns during collaboration between human workers and robots.

Gesture-based communication provides a fast and direct means for the human worker to convey simple and fast control commands and responses to the robotic companion, making it a useful and naturalistic way of communicating in the collaborative environment of the aCell. The Interaction Manager integrates eight different communication signals, providing a rich set of gestures for effective communication between the worker and the robotic system.

aCell-to-human communication is enabled through a tablet-based application, representing the main communication channel through which the system can provide detailed information about the current task’s state. Additionally, through this Interaction Manager application (IM app), the system can request human intervention during the disassembly or request human input for problem-solving or decision-making when unforeseen problematic situations are faced.

These two main modes of interaction have been chosen to cover the speed-accuracy trade-off, with gestures for simpler but time-sensitive interactions, and tablets for slower but more fine-grained information exchange. This dual communication paradigm accommodates individual human preferences and ensures efficient collaboration in various human-robot collaboration scenarios, as we will see in the Results section.

The Interaction Learner adds a level of personalization on top of the Interaction Manager functionalities by providing it with an adaptive mechanism to support human-robot decision-making based on the prior history of interactions between the human and the cobot. Its main function is to keep track of human-robot interactions and human feedback during error-handling scenarios. It computes useful statistics based on the history of human-robot interactions, and when a similar situation is encountered, it adapts the options displayed in the tablet by the Interaction Manager in a way that enhances collaborative decision-making by highlighting on the menu the most frequently selected options by that worker in a given situation. This level of adaptation takes into account the human-robot interactions at each specific step during the disassembly and for each worker in particular.

The design of the Socially Adaptive Safety Engine (SASE) incorporates a set of reactive control systems inspired by Pavlovian appetitive and aversive drives Freire et al. (2020a). This approach shapes the SASE’s functionality, guiding its interactions in the DAC-HRC architecture to align with principles of both harm avoidance and proactive cooperation. This incorporation allows the Socially Adaptive Safety Engine to adapt its behavior dynamically, not only avoiding harm but also optimizing operation parameters such as speed, distance, and task allocation based on the unique context of each human worker. The Worker Model, integral to the contextual layer of the DAC-HRC architecture, helps personalize the interaction, treating each worker as a distinct individual with specific preferences and needs.

The Socially Adaptive Safety Engine module, in charge of providing a context-aware and personalized safety control system, spans across three layers of the DAC-HRC architecture. The SASE’s reactive layer integrates several homeostatic modules whose purpose is to monitor key aspects of the human-robot interaction. The goal of each homeostatic module is twofold: to keep its desired variable within the optimal range of operation, and to exert control when that variable trespasses the safe range. The current implementation comprises the homeostatic control of key proxemics variables in HRI, such as the human-robot interaction security distance, the robot movement speed, and the robot action execution speed. When any of those variables reach or trespass their threshold, the control response can be either a direct modification of the exceeded value -in the case of speed modulation-, or a command directed to stop the robot’s current action -in case the HRI distances are trespassed. For instance, if the actual detected distance between the human and the robot is below the desired safety value, the homeostatic control system will generate a stop signal and the robot will not move until the actual distance goes back to the desired range.

The Socially Adaptive Safety Engine’s adaptive layer is composed of the allostatic control module. This is the key mechanism by which the SASE can adapt the interaction of the robot to its changing environment. This module is in charge of the transformation of the environmental information provided by the contextual layer and modifying the desired parameters of the subsequent homeostatic regulatory mechanisms of the reactive layer. For instance, when the robot is handling a dangerous tool, the allostatic control module gets this information and adapts the desired safety HRI distance, as well as the speed at which the robot will operate when being close to a human.

The Socially Adaptive Safety Engine’s contextual layer deals with the integration of the relevant environmental, social, and material information that comes from other modules of the DAC-HRC architecture. It endows the SASE with context awareness. The constant integration of these different sources of information defines the specific context at every point in time, thus allowing the SASE to monitor and adapt the behavior of the robot to the changing conditions of its surroundings. For instance, the contextual layer can obtain information in real-time about the HRI preferences of the currently detected human worker, the risk level of the current robot action, and the information about the current tool being used by the robot (if any).

The incorporation of the reactive control systems inspired by the Pavlovian appetitive and aversive drives allows the Socially Adaptive Safety Engine to adapt its behavior dynamically, not only avoiding harm but also optimizing operation parameters such as speed, distance, and task allocation based on the unique context of each human worker. The worker model, integral to the contextual layer of the DAC-HRC architecture, helps personalize the interaction, treating each worker as a distinct individual with specific preferences and needs.

The Worker Model is composed of short-term and long-term memory buffers along with its reactive and adaptive input processing layers. The Worker Model’s reactive layer serves as a first data integration step, gathering information from several input sources, whereas its adaptive layer processes the raw data in order to produce new parameters that will be used by other modules of the Worker Model and the DAC-HRC architecture. The online information gathered by the Worker Model’s reactive layer is transiently stored in the short-term memory buffer before it is further processed by the adaptive Layer to generate new relevant information about the worker and their interaction with the system. For instance, the short-term memory can store the timings of past interactions during a disassembly step, while the adaptive layer generates estimates of current task duration based on this input. The type of input information gathered by the Worker Model can be divided into offline and online variables:

•Offline variables - This type of data is mostly static, as it will not vary throughout the session (e.g., age, gender, language, and interaction preferences). This information is acquired through preliminary questionnaires before engaging with the system and defines the profile of each user based on demographic information and her opinion towards robots.

The technical implementation of the Worker Model is based on two main components: the Worker Model’s API and the Worker Model Database. The database implements the long-term memory component of the Worker Model. Its function is to store all the information related to each worker and to keep it up to date. It is deployed as a document-oriented database using MongoDB ¹ , where each worker profile is stored as a unique document. Each worker model entry is initialized with the offline variables acquired from the worker profile and questionnaires. Additionally, it also stores the main statistics of each interaction between the worker and the DAC-HRC collaborative architecture that has been extracted by the Worker Model API, such as the expected task duration or the history of interactions with the tablet.

All the communications with the database are centrally controlled by the Worker Model’s API, which integrates the reactive and adaptive input processing layers along with the short-term memory component of the Worker Model. The API’s function is twofold: it performs the basic CRUD (create, read, update, and delete) operations that keep the database up to date, and it is in charge of filtering the online and state variables to produce the task- and interaction-relevant outputs of the Worker Model. The API is written in Python and communicates with the database using BSON as the data interchange format.