Introduction and background

Trust is essential in shaping human-human relationships [1]. As autonomous agents are starting to enter our everyday environments, we see an increasing number of researchers in Human-Robot Interaction (HRI) directing their efforts towards understanding how this factor influences the relationships between humans and intelligent machines [2]. For example, in a collaborative setting between humans and robots, establishing a trust relationship enables the user to delegate a portion of the shared task to the robot [3,4,5]. As a result, the user can concentrate on their own responsibilities within the task, thus enhancing the overall outcome [6]. Trust in HRI has been defined as the "attitude that an agent will help achieve an individual's goals in a situation characterised by uncertainty and vulnerability" [7], and it is mainly studied from the point of view of the human. In fact, most efforts have been directed towards understanding what robotic characteristics facilitate or hinder human partners' trust [8,9], or what strategies lead to trust repair once the latter is lost [10]. Research on trust and trustworthiness in automation has become even more critical with the integration of Artificial Intelligence (AI) in the decision-making processes of autonomous agents.

Within this research panorama, we are witnessing an increasing interest in trying to mathematically model human trust towards robots in an attempt to understand and exploit the human partner's internal state [2]. We term this modeling attempt as "Natural Trust". The focus of the latter is understanding the human partner's internal model of the robot to facilitate the establishment of trust. In a dual fashion, few mathematical models address the trust that an artificial agent could have towards the user it is interacting with, named "Artificial Trust" [11,12]. Both these forms of trust are rarely utilized during a human-robot interaction to alter the behavioral policy of the artificial agent.

From these considerations, a gap in the current state-of-the-art emerges: robots are often considered as passive receptacles of trust, and not as active social agents that makes use of such trust to improve their own behaviour. Starting from the consideration that trust is a bidirectional relationship needed to successfully complete collaborative tasks [13], we are interested in investigating when and how it is possible for robots to model the trustworthiness of their human partners [14] and/or the trust put by the users in them, to allow them to exploit this knowledge and enhance their interactions with the users. Attempts at regularizing the design of models of trust for robots towards their users or other agents are still scarce, yet of great importance. Research in the current literature [15,14] has shown that robots that possess cognitive mechanisms to identify and anticipate mistakes or deceptions in their human partner's strategy (in other words, evaluating their trustworthiness) can increase the success rate of joint collaborative tasks.

In an attempt to address this knowledge gap, we propose the term "Computational Trust" (CT) to refer to the mathematical models that can be used by a robot or an artificial agent to perform trust evaluations on other agents. This term incorporates both cases of Artificial Trust and Natural Trust. We discuss initial results from a systematic review we are currently finalizing. This work aims to be the first systematic dive into this new research domain.

Discussion and preliminary results

Figure 1: Keyword network of the selected set of papers (generated through VOSviewer [16]).

From the analysis of the state-of-the-art in trust in HRI detailed in Section 1, we have identified a gap in the current research landscape. Specifically, many works presented in the literature focus on the human side of robotic trust, i.e., the trust that humans place in robots. As we have seen, this is an important issue to consider during the worldwide effort to integrate social robots into our daily lives and environments. Despite this, we argue that this does not draw a complete picture of trust relationships between humanity and automation. In fact, there is evidence that trust is a bidirectional relationship and that the dynamics of mutual trust should not be ignored [13].

To address this gap in the current knowledge, we have performed a systematic review of the current literature, searching for scientific publications that describe computational models of CT in HRI. Our investigation across three major scientific databases (IEEE Xplore, Scopus, and Web of Science) has led to the selection of 101 papers. By analysing them, we have generated a map of the co-occurrence of their keywords, reported in Figure 1. In its upper branch, the keyword "robot trust" appears, which is very closely linked to CT. Not only it appears in very recent publications, but it derives from topics such as "decision making" and "cognitive model". This is evidence of a recent trend in the literature to embed trust within robotic agents. However, this term is ambiguous, as it is mainly used to depict the classical perspective of human trust in robots. To avoid any confusion, we should instead refer to the more objective term of CT, which unambiguously refers to any way of mathematically modeling trust estimates in HRI. Furthermore, the keyword "human-robot collaboration" is closely linked to "trust". Finally, it is worth noting that, since we looked for CT models, it emerges that these models are increasingly being used to modulate the behaviour of robots during collaborative tasks with humans. All this evidence underscores the importance of pursuing standards in designing CT models.