Robots are envisioned to be able to process many complex inputs from the environment and be active participants in many aspects of life, including work environments, home assistance, battlefield and crisis response, and others. Therefore, robots are envisioned to transition from tool to teammate as humans transition from operator to teammate in an interaction more akin to human-human teamwork. These envisioned transitions raise a number of general questions: How would human interaction with the robot be affected? How would performance of the human-robot team be affected? How would human performance or behavior be affected? Although there are numerous tasks, environments, and situations of human-robot collaboration, in order to best clarify the role of trust we distinguish two general types of interactions of humans and robots: performance-based interactions, where the focus is on the human influencing/controlling the robot so it can perform useful tasks for the human, and social-based interactions, where the focus is on how the robot’s behavior influences the human’s beliefs and behavior. In both these cases, the human is the trustor and the robot the trustee. In particular, in performance based interactions there is a particular task with a clear performance goal. An example of performance-based interactions is where human and robot collaborate in manufacturing assembly, or a UAV performing surveillance and recognition of victims in a search and rescue mission. Here measures of performance could be accuracy and timing to complete the task. On the other hand, in social interactions, the performance goal is not as crisply defined. An example of such a task is the ability of a robot to influence a human to reveal private knowledge, or how a robot can influence a human to take medicine or do useful exercises.
A large body of HRI research investigating factors thought to affect behavior via trust, such as reliability, rely strictly on behavioral measures without reference to trust. Meyer’s [
82] expected value (EV) theory of alarms provides one alternative by describing the human’s choice as one between compliance (responding to an alarm) and reliance (not responding in the absence of an alarm). The expected values of these decisions are determined by the utilities associated with an uncorrected fault, the cost of intervention and the probabilities of misses (affecting reliance) and false alarms (affecting compliance). Research in [
31], for example, investigated the effects of unmanned aerial vehicle (UAV) false alarms and misses on operator reliance inferred from longer reaction times for misses and compliance inferred from shorter reaction times to alarms. While reliance/compliance effects were not found, higher false alarm rates correlated with poorer performance on a monitoring task, while misses correlated with poorer performance on a parallel inspection task. A similar study by [
20] of unmanned ground vehicle (UGV) control found participants with higher perceived attentional control were more adversely affected by false alarms (under-compliance) while those with low perceived attentional control were more strongly affected by misses (over-reliance). Reliance and compliance can be measured in much the same way for homogeneous teams of robots as illustrated by a follow up study of teams of UGVs [
19] of similar design and results. A similar study [
26] involved multiple UAVs manipulating ATR reliability and administering a trust questionnaire, again finding that ratings of trust increased with reliability.
Transparency, common ground, or shared mental models involve a second construct (“process” [
58] or “integrity” [
76]) believed to affect trust. According to these models, the extent to which a human can understand the way in which an autonomous system works and predict its behavior will influence trust in the system. There is far less research on effects of transparency, with most involving level of automation manipulations. An early study [
60] in which all conditions received full information found best performance for an intermediate level of automation that facilitated checks of accuracy (was transparent). Participants, however, made substantially greater use of a higher level of automation that provided an opaque recommendation. In this study, ratings of trust were affected by reliability but not transparency. More recent studies have equated transparency with additional information providing insight into robot behavior. Researchers in [
9] compared conditions in which participants observed a simulated robot represented on a map by a status icon (level of transparency 1), overlaid with environmental information such as terrain (level 2), or with additional uncertainty and projection information (level 3). Note that these levels are distinct from Sheridan’s Levels of Automation mentioned previously. What might appear as erratic behavior in level 1, for example, might be “explained”’ by the terrain being navigated in level 2. Participant’s ratings of trust were higher for levels 2 and 3. A second study manipulated transparency by comparing minimal (such as static image) contextual (such as video clip) and constant (such as video) information for a simulated robot team mate with which participants had intermittent interactions but found no significant differences in trust. In [
126], researchers took a different approach to transparency by having a simulated robot provide “explanations” of its actions. The robot guided by a POMDP model can make different aspects of its decision making such as beliefs (probability of dangerous chemicals in building) or capabilities (ATR has 70% reliability) available to its human partner. Robot reliability affected both performance and trust. Explanations did not improve performance but did increase trust among those in the high reliability condition. As these studies suggest, reliability appears to have a large effect on trust, reliance/compliance, and performance, while transparency about function has a relatively minor one, primarily influencing trust. The third component of trust in robot’s “purpose” [
58] or “benevolence” [
76] has been attributed [
69,
70,
95] to “transparency” as conveyed by appearance discussed in Sect.
8.6.2. By this interpretation, matching human expectations aroused by a robot’s appearance to its purpose and capabilities can make interactions more transparent by providing a more accurate model to the human.
Studies discussed to this point have treated trust as a dependent variable to be measured at the end of a trial and have investigated whether or not it had been affected by characteristics of the robot or situation. If trust of a robot is modified through a process of interaction, however, it must be continuously varying as evidence accumulates of its trustworthiness or untrustworthiness. This was precisely the conception of trust investigated by Lee and Moray [
56] in their seminal study but has been infrequently employed since. An recent example of such a study is reported in [
29] where a series of experiments addressing temporal aspects of trust involving levels of automation and robot reliability have been conducted using a robot navigation and barriers task. In that task, a robot navigates through a course of boxes with labels that the operator can read through the robot’s camera and QR codes presumed readable by the robot. The labels contain directions such as “turn right” or “U turn”. In automation modes, robots follow a predetermined course with “failures” appearing to be misread QR codes. Operators can choose either the automation mode or a manual mode in which they determine the direction the robot takes. An initial experiment [
29] investigated the effects of reliability drops at different intervals across a trial, finding that decline in trust as measured by post trial survey was greatest if the reliability decline occurred in the middle or final segments. In subsequent experiments, trust ratings were collected continuously by periodic button presses indicating increase or decrease in trust. These studies [
30,
49] confirmed the primacy-recency bias
in episodes of unreliability and the contribution of transparency in the form of confidence feedback from the robot.
Work in [
24] collected similar periodic measures of trust using brief periodically presented questionnaires to participants performing a multi-UAV supervision task to test effects of priming on trust. These same data were used to fit a model similar to that formalized by [
39] using decision field theory to address the decision to rely on the automation/robot’s capabilities or to manually intervene based on the balance between the operator’s self-confidence and her trust in the automation/robot. The model contains parameters characterizing information conveyed to operator, inertia in changing beliefs, noise, uncertainty, growth-decay rates for trust and self-confidence, and an inhibitory threshold for shifting between responses. By fitting these parameters to human subject data, the time course of trust (as defined by the model) can be inferred. An additional study of UAV control [
38] has also demonstrated good fits for dynamic trust models with matches within 2.3% for control over teams of UGVs. By predicting effects of reliability and initial trust on system performance, such models might be used to select appropriate levels of automation or provide feedback to human operators. In another study involving assisted driving [
123], the researchers use both objective (car position, velocity, acceleration, and lane marking scanners) and subjective (gaze detection and foot location) to train a mathematical model to recognize and diagnose over-reliance on the automation. The authors show that their models can be applied to other domains outside automation-assisted driving as well.
Willingness to rely on the automation has been found in the automation literature to correlate with user’s self-confidence in their ability to perform the task [
57]. It has been found that if a user is more confident in their own ability to perform the task, they will take control of the automation more frequently if they perceive that the automation does not perform well. However, as robots are envisioned to be deployed in increasingly risky situations, it may be the case that a user (e.g. a soldier) may elect to use a robot for bomb disposal irrespective of his confidence in performing the task. Another factor that has considerably influenced use of automation is user workload. It has been found in the literature that users exhibit over-reliance [
7,
40] on the automation in high workload conditions.
Experiments in [
104] show that people over-trusted a robot in fire emergency evacuation scenarios conducted with a real robot in a campus building, although the robot was shown to be defective in various ways (e.g. taking a circuitous route rather then the efficient route in guiding the participant in a waiting room before the emergency started). It was hypothesized by the experimenters that the participants, having experienced an interaction with a defective robot, would decrease their trust (as opposed to a non-defective robot), and also that participants’ self-reported trust would correlate with their behavior (i.e their decision to follow the robot or not). The results showed that, in general, participants did not rate the non-efficient robot as a bad guide, and even the ones that rated it poorly still followed it during the emergency. In other words, trust rating and trust behavior were not correlated. Interestingly enough, participants in a previous study with similar scenarios of emergency evacuation
in simulation by the same researchers [
103] behaved differently, namely participants rated less reliant simulated robots as less trustworthy and were less prone to follow them in the evacuation. The results from the simulation studies of emergency evacuation, namely positive correlation between participants’ trust assessment and behavior, are similar to results in low risk studies [
30]. These contradictory results point strongly that more research needs to be done to refine robot, operator and task-context variables and relations that would lead to correct trust calibration, and better understanding of the relationship between trust and performance in human robot interaction.
One important issue is how an agent forms trust in agents it has not encountered before. One approach from the literature in multiagent systems (MAS) investigates how trust forms in ad hoc groups, where agents that had not interacted before come together for short periods of time to interact and achieve a goal, after which they disband. In such scenarios, a decision tree model based on both trust and other factors (such as incentives and reputation) can be used [
13]. A significant problem in such systems, known as the
cold start problem, is that when such groups form there is little to no prior information on which to base trust assessments. In other words, how does an agent choose who to trust and interact with when they have no information on any agent? Recent work has focused on bootstrapping such trust assessments by using stereotypes [
12]. Similar to stereotypes used in interpersonal interactions among humans, stereotypes in MAS are quick judgements based on easily observable features of the other agent. However, whereby human judgements are often clouded by cultural or societal biases
, stereotypes in MAS can be constructed in a way that maximizes the accuracy. Further work by the researchers in [
14] shows how stereotypes in MAS can be spread throughout the group to improve others’ trust assessments, and can be used by agents to detect unwanted biases received from others in the group. In [
15], the authors show how this work can be used by organizations to create decision models based on trust assessments from stereotypes and other historical information about the other agents.
8.6.1.1 Towards Co-adaptive Trust
In other studies [
129,
130], Xu and Dudek create an online trust model to allow a robot or other automation to assess the operator’s trust in the system while a mission is ongoing, using the results of the model to adjust the automation behavior on the fly to adapt to the estimated trust level. Their end goal is
trust-seeking adaptive robots, which seek to actively monitor and adapt to the estimated trust of the user to allow for greater efficiency in human-robot interactions. Importantly, the authors combined common objective, yet indirect, measures of trust (such as quantity and type of user interaction), with a subjective measure in the form of periodical queries to the operator about their current degree of trust.
In an attempt to develop an objective and direct measure of trust the human has in the system, the authors of [
36] use a mathematical decision model to estimate trust by determining the expected value of decisions a trusting operator would make, and then evaluate the user’s decisions in relation to this model. In other words, if the operator deviates largely from the expected value of their decisions, they are said to be less trusting, and vice versa. In another study [
108], the authors use two-way trust to adjust the relative contribution of the human input to that of the autonomous controller, as well as the haptic feedback provided to the human operator. They model both robot-to-human and human-to-robot trust, with lower values of the former triggering higher levels of force feedback, and lower values of the latter triggering a higher degree of human control over that of the autonomous robot controller. The authors demonstrate their model can significantly improve performance and lower the workload of operators when compared to previous models and manual control only.
These studies help introduce the idea of “inverse trust”. The inverse trust problem is defined in [
34] as determining how “an autonomous agent can modify it’s behavior in an attempt to increase the trust a human operator will have in it”. In this paper, the authors base this measure largely on the number of times the automation is interrupted by a human operator, and uses this to evaluate the autonomous agent’s assessment of change in the operator’s trust level. Instead of determining an absolute numerical value of trust, the authors choose to have the automation estimate
changes in the human’s trust level. This is followed in [
35] by studies in simulation validating their inverse trust model.
8.6.2 Social-Based Interactions: Robots Influencing Humans
Social robotics deals with humans and robots interacting in ways humans typically interact with each other. In most of these studies, the robot—either by its appearance or its behavior—influences the human’s beliefs about trustworthiness, feelings of companionship, comfort, feelings of connectedness with the robot, or behavior (such as whether the human discloses secrets to the robot or follows the robot’s recommendations). This is distinct from the prior work discussed, such as ATR, where a robot’s actions are not typically meant to influence the feelings or behaviors of its operator. These social human-robot interactions contain affective elements that are closer to human-human interactions. There is a body of literature that looked at how robot characteristics affected ratings of animacy and other human-like characteristics, as well as trust in the robot, without explicitly naming a performance or social goal that the robot would perform. It has been consistently found in the social robotics literature that people tend to judge robot characteristics, such as reliability and intelligence, based on robot appearance. For example, people ascribe human qualities to robots that look more anthropomorphic. Another result of people’s tendency to anthropomorphize robots is that they tend to ascribe animacy and intent to robots. This finding has not been reported just for robots [
109] but even for simple moving shapes [
44,
48]. Kiesler and Goetz [
52] found that people rated more anthropomorphic looking robots as more reliable. Castro-Gonzalez et al. [
18] investigated how the combination of movement characteristics with body appearance can influence people’s attributions of animacy, liekeability, trustworthiness, and unpleasantness. They found that naturalistic motion was judged to be more animate, but only if the robot had a human appearance. Moreover, naturalistic motion improved ratings of likeability irrespective of the robot’s appearance. More interestingly, a robot with human-like appearance was rated as more disturbing when its movements were more naturalistic. Participants also ascribe personality traits to robots based on appearance. For instance, in [
118], robots with spider legs were rated as more aggressive whereas robots with arms rated as more intelligent than those without arms. Physical appearance is not the only attribute that influences human judgment about robot intelligence and knowledge. For example, [
59] found that robots that spoke a particular language (e.g. Chinese) were rated higher in their purported knowledge of Chinese landmarks than robots that spoke English.
Robot appearance, physical presence [
3], and matched speech [
94] are likely to engender trust in the robot [
124] found that empathetic language and physical expression elicits higher trust [
62] found that highly expressive pedagogical interfaces engender more trust. A recent meta-analysis by Hancock et al. [
43] found that robot characteristics such as reliability, behaviors and transparency influenced people’s rating of trust in a robot. Besides these characteristics, the researchers in [
43] also found that anthropomorphic qualities also had a strong influence on ratings of trust, and that trust in robots is influenced by experience with the robot.
Martelato et al. [
73] found that if the robot is more expressive, this encourages participants to disclose information about themselves. However, counter to their hypotheses, disclosure of private information by the robot, a behavior that the authors labelled as making the robot more vulnerable, did not engender increased willingness to disclose on the part of the participants. In a study on willingness of children to disclose secrets, Bethel et al. [
5] found in a qualitative study that preschool children were found to be as likely to share a secret with an adult as with a humanoid robot.
An interesting study is reported in [
111], where the authors studied how errors performed by the robot affect human trustworthiness and willingness of the human to subsequently comply with the robot’s (somewhat unusual) requests. Participants interacted with a home companion robot, in the experimental room that was the pretend home of the robot’s human owner in two conditions, (a) where the robot did not make mistakes and (b) where the robot made mistakes. The study found that the participants’ assessment of robot reliability and trustworthiness was decreased significantly in the faulty robot condition; nevertheless, the participants were not substantially influence in their decisions to comply with the robot’s unusual requests. It was further found that the nature of the request (revocable versus irrevocable) influenced the participants’ decisions on compliance. Interestingly, the results in this study also show that participants attributed less anthropomorphism when the robot made errors, which contradict those found by an earlier study the same authors had performed [
110].