Human-robot interaction nalized to cooperation and teamwork is a demanding research task, both under the development and the implementation point of view. In this context, cognitive architectures are a useful means for representing the cognitive perception-action cycle leading the decision-making process. In this paper, we present ongoing work on a cognitive architecture whose modules consider the possibility to represent the decision-making process starting from the observation of the environment and also of the inner world, populated by trust attitudes, emotions, capabilities and so on, and the world of the other in the environment.
A human-robot interaction system is created for di erent purposes, from assisting in home, healthcare or learning to safety or search and rescue scenarios and so on; in any case, it is a complex system to design and develop. If also, we consider, a human-robot team where the robot and the human have to autonomously cooperate, communicate and collaborate to reach a common and shared objective, we are talking of a more complex system. Here the term \complex" is used to mean that, in such a context, the robot, as well as the whole system, does not have a behavior that can be analyzed and implemented as a sum of the parts because it assumes an emergence of behaviors at runtime. For instance, let us suppose to have a team made up of a human and a robot who has to carry out a task, known to both of them, interacting with each other. During the design phase, all the actions that each of the two must perform and all the communications that they must exchange during the interaction can be set. During the execution phase, if the task is to be carried out in a dynamic environment, the interactions between the robot and the human and between them and the environment inevitably change the state of the world that can thus provide new constraints or requirements for achieving the initial goal. The actual behavior of the whole system comes out at runtime and the team members need to be able to respond to changes e ciently.
The most signi cant di culties in these types of systems are related to equipping the robot with the ability to select at runtime the best action to perform to achieve the team's goal. These types of problems are often studied and implemented by looking at their human counterpart, in our case the human-human team. The question is: how does a human being act in a team whose goals he knows and shares in a changing environment? And, how can this be reported in a human-robot team? Usually, a human being grounds his decisional process on a set of factors, such as knowledge of the surrounding environment and knowledge of a set of possible plans allowing to achieve a goal.
These factors, which could be described as reactive, occur when the situation being addressed is known and the changes are foreseeable before an action plan is established. When one is engaged in a cooperative and dynamic context some psychological factors intervene, that is, the whole set of internal states that trigger a decision and are closely related to the mental state of the human being, the knowledge of himself and the other elements of the team. For example, one element of a team may understand to be not able to perform a certain action that can lead to the achievement of the common goal. It cannot, also, want to do an action, and delegate it to the other team member based on a certain level of trust in the skills of the other. Or it could be in such an emotional state (for example euphoria) to want to do more than what was told during the design phase and so propose to the other component to carry out some actions or even decide to act independently on the behalf of the other.
In this context, the human behaves, in all respects, as a cognitive agent so a research eld that is spreading nowadays is that of cognitive architectures. In recent years, various cognitive architectures have been studied and have served to examine how human beings behave, but they have also been adopted as the basis for the implementation of robotic systems that act like humans.
This paper presents an ongoing study focused on the use of self-modeling and the theory of mind in order to propose a cognitive architecture. It would allow modeling a human-robot teaming system that cooperates to achieve a common goal and, as well as a human team, applies a decision-making process that is driven from the objective situation of the environment, but also from the knowledge that each cognitive agent has of itself and the other members of the team.
In the rest of the paper, we set state of the art in the cognitive architecture of human-robot interaction systems and the motivation of our work, then we present our proposal for a cognitive architecture and nally we draw some discussions and conclusions. 2
Modeling and understanding cognitive processes are the primary aim of the research on cognitive architectures. A cognitive architecture determines the structure of a computational cognitive model, applied in a generic domain, by underlying the infrastructure of an intelligent system.
Typically, cognitive architectures include modules representing the short and long-term memory and the functional processes operating on memory modules for realizing learning and action mechanisms. Several cognitive architectures have been proposed in the literature; our work starts from the analysis of a lot of them in order to deduce the elements suitable for our application context and whether and how they needed to be extended and integrated. We have been driven by the assumption that cognitive architectures model intelligent systems where components' behaviors are not established and coded at design time but arise from perception and knowledge.
Among the most known cognitive architectures, we may count those inspired
by psychology and biology (for instance CLARION, SOAR, ACT-R . . . ) and
those by agent technology, like for instance LIDA [
select
In addition to these two modules, there is a basic module that manages
everything related to data storage for further processing. This allows to represent,
and therefore to implement, a system in which the decisional process passes from
the elaboration of a set of data, relative to everything that is perceived,
memorized in the system. In existing architectures, the memory is mainly divided into
short-term memory and long-term memory. The short-term memory stores all
the facts and elements relating to the current moment. The long-term memory
contains all data and events that are held inde nitely. Long-term memory is
further divided into explicit and implicit memory in all its di erent forms (episodic,
semantic, autobiographical...) already widely debated in the literature [
So, the memory module plays a fundamental role in reasoning on and learning the current situation in order to trigger the decision process.
Nevertheless, frequently human-robot teaming applications are not so simple to be sketched through these three simple modules but it is necessary, at least, that the memory mechanism also refers to all the evidence coming from perception and from knowledge of self and from the perception and knowledge of all the other agents present in the environment. This means that, in a human-robot teaming context, architectures must contain the modules useful for the representation of itself, of the (physical) world around and of the other, including all the mental states that this fact entails.
To date, most architectures base their decision and learning process only on
the concept of stored data or fact and not on the notion of mental state, and also
the implementation part of most known cognitive architectures is still missing.
Our contribution lies in the creation of a cognitive architecture in which memory
also contains all the information about the mental state so that the perceive-act
cycle becomes what we call the perceive-proact cycle. Besides, the architecture
we propose can be easily mapped to a BDI agent architecture [
In the next paragraph, we explain how the architecture we propose includes the modules to model and represent a human-robot teaming system. 3
In Fig. 2 we illustrate the cognitive architecture we propose for realizing humanrobot teaming interaction. The architecture is composed of the several known modules from general cognitive architecture and is enriched with the modules devoted to employing decision functionalities.
Our starting point is the generalization of all the characteristics of the cognitive architectures studied that led, together with the in-depth analysis of LIDA, to the representation given in Fig. 1. The gure shows what has been said above regarding a perception-action cycle; the workspace is the module used for interaction with the work environment.
In LIDA the workspace module takes into account two di erent facts: the changing environment, in the cognitive cycle the agents are able to make an internal representation of the world; all this helps agents in selecting the actions to be done also taking into account the conscious part of the agent. Consciousness in LIDA is the mental counterpart of attention, that is, of the perception process. LIDA gives a complete representation of a cognitive cycle but it is not suitable to represent interactions in teaming because it lacks the part related to the representation of the self and to the representation of the other.
Before going into the details of the proposed architecture, it is worth to make a hint to some necessary starting hypotheses. As was said in the introduction, a human-robot system working in team to achieve a common objective is a complex system. A complex system, by de nition, is a system made up of complex components that interact with each other and with the surrounding environment. The global behavior and characteristics of the system emerge from interactions. The behavior of a complex system cannot be analyzed and implemented as the sum of the single components, the system must be seen as a whole. In our case, we consider the system as composed of a set of cognitive agents, living in an environment made by inanimate objects and cognitive agents at the same time. Analysis and design of such a system must begin by choosing the perspective from which we look at the system.
For example, if the perspective is that of the cognitive agent (the robot), it will see the system as the set of inanimate objects on which he can act. It will see itself with everything he has inside (his goals, his capabilities, his mental states) and he will see the other (any other cognitive agent) with everything he has inside and with the set of objects on which the other can act. So the man-robot system is, to all intents and purposes, a system of systems in which the environment is not external like something with which he only interacts but is an integral part of the system itself.
This is the main di erence between our architecture and the existing ones.
Such a feature allows us to represent and then implement a decision-making
process based on a series of factors similar to human ones, such as the sense
of self, elements of the theory of the mind, trust, emotions and everything can
generate a mental state. In so doing we are able to realize a change of design
paradigm, from design time to runtime, already studied and analyzed in some
previous works [
Action Selection
Fig. 2 our architecture is shown. As it can be noted, the cycle Monitor,
Analyzing, Planning and Executing (MAPE), which is the basis of all the
implementations of complex autonomous systems [
In [
The Decision Process part of the architecture is centered on Anticipation and Motivation. According to our architecture, the robot does not act only after the reasoning process based on a certain data stored in the memory, but also and mainly after evaluating the anticipation of its actions. During the anticipation process, the current situation is generated; it represents the state of the world corresponding to the currently selected action. The current situation is elaborated on the basis of motivations, goals and all those elements are in the memory thus getting the execution launched. A queue of possible situations is also created, intended as a set of pre-conditions, objectives and knowledge to achieve them, and postconditions on the objectives. The robot can draw on all these elements at any time to respond to changes and still maintain its initial target. The Motivation module is the one triggering the anticipation and the action selection. It is the core of the decision process, here all the information and process for elaborating mental states reside and it is the module devoted to the representation of inner and word of each cognitive agent. Through this module, therefore, it is possible to make decisions about the actions being conveyed by the sense of self, by the ability to attribute mental states (belief, desire, intention, knowledge, capabilities) to oneself and to others and by the understanding that others have di erent mental states, by emotions, by the level of trust in the abilities (or more generally by trust) of others and of oneself.
The experiments 3 we are performing in our laboratory are con rming that with these modules, the architecture we propose can be easily mapped into a BDI agent architecture thus giving the possibility to implement a human-robot interaction system with a multi-agent system made of cognitive agents. 3 Showing the experiments is not in the scope of this paper.
Designing and implementing a human-robot interaction system in a dynamic environment and for tasks that require team organization is a challenging issue. The literature provides a rich set of cognitive architectures that allow modeling the human-robot interactions through the classical perception-action cycle of a cognitive agent. However, the application domain we are working on is such that considering and using the action-perception cycle only is not enough and also, most existing cognitive architectures do not provide a usefully implemented counterpart. In this work, we have therefore generalized and extended the existing architectures in order to obtain the modules that meet our requirements, namely to create a human-robot interaction on the basis of the human-human interaction in a team that cooperates in the same highly evolving environment to achieve a known and shared goal.
The main contribution of this work is the module for the anticipation of the situation and the one for the representation of the internal state of the robot and for the representation of knowledge about the internal state of the other. In this way, we can integrate the use of self-modeling and the theory of mind in team interaction and besides, we can scale the use of this module including a whole set of mental states that are typical in the human, therefore emotions, levels of trust in oneself and the others, etc. It was also fundamental to make a change in the abstract representation of the environment, from the element with which the robot interacts, as a portion of the world outside, to the element that is the world together with the robot itself as in a sort of recursive structure.
In this way, for example, the work of the execution module on the environment and therefore on any elements of the environment can be con gured, for example, in the delegation an action from the robot to the human being when it cannot do an action for any reason generated by the knowledge of himself and the environment.
Finally, the module we have called Motivation, allows us to verticalize the architecture from a theoretical level to a system level with a mapping one by one with a BDI agent architecture. We are already doing several experiments on this and in the future, we will realize the whole system in which the internal state of the robot completely conveys decisions on actions. We expect, for example, that a robot at the beginning of its activities, which has learned few elements of the surrounding world and of the human being with whom it is interacting, has little margin for autonomous choices on the actions to be taken and is only able to implement those given at design time. After a certain number of interactions, it will have acquired knowledge, a certain level of trust in the other and be able to act autonomously.
Acknowledgment. This material is based upon work supported by the Air Force O ce of Scienti c Research under award number FA9550-17-1-0232.