Robot’s errors

People’s perception of a robot’s reliability depends not only on the ability of a robot to complete a task, but also by its behaviours to reach a goal. Several studies (Honig et al.2018; Lemaignan et al.2015; Short et al.2010) showed that people might consider unexpected and incoherent behaviours, perceived failures, and actual failures as robot errors. Humans may perceive erroneous robot behaviours according to their expectations of a robot’s proper functions (Michael L. Walters et al.2011). For example, a robot that navigates too slowly might be considered having faulty behaviours. Honing et al. (2018) proposed a taxonomy for classifying possible types of robotic failures, also referred as faults (see Figure 2.1). They identified two principal categories of errors: technical failures and interaction failures. Technical failures are considered errors caused by hardware or software problems, which can depend on an erroneous design, communication or processing. In contrast, interaction errors are related to social norm violations, organisational and mental-models faults in the interaction with the environment and other agents, including people.

Bainbridge et al. (2011) found that participants were happy to follow a robot’s instructions to throw books in the trash if the robot was present in the room with them, but not when the robot was not physically in the same room.

Other studies (M. Desai et al. 2012; Munjal Desai et al.2013) showed that the order of presentation of the decreased reliability produces an evident drop in the trust in the robot which can be restored by continuing the interaction. They showed that warning the participants about a drop in the robot’s performance can mitigate the loss in trust.

Wang et al. (2015)’s studies showed that the frequency and significance of errors can impact humans’ trust in an imperfect on-line system. They showed that people are not willing to follow an imperfect robot if the outcomes are severe, and could lead to permanent consequences, e.g. death of the human.

Booth et al. (2017) investigated participants’ responses to a robot’s request to move in a secure-access student dormitory. They conducted the experiment with two conditions: 1) an anonymous robot and 2) a food delivery robot, where both asked to enter the building. They observed that participants were more likely to let the food delivery robot enter the building or in situations when they were in a group.

Figure 2.1: A human-robot failure taxonomy reproduced by Honig et al. (2018).

Robinette et al. (2015) investigated the effects of apologies, promises and addi- tional reasons given by a robot for its errors on participants’ trust in a simulated fire evacuation scenario conducted in a virtual environment. They showed that participants’ trust was repaired if the robot apologised and promised to not repeat the error soon after it made the error but not during the emergency.

Salem et al. (2015) studied human perception of trust in robots, and how willing they are to follow a robot showing faulty behaviours. They showed that no matter how erratic the behaviour of the robots, participants followed the instructions of the robots. Similarly, Robinette et al. (2016) used an emergency evacuation scenario, with artificial smoke and a smoke detector, in which a robot guided a person to an exit, in order to study how willing were humans to follow a robot that had previously exhibit erratic behaviour. Their results indicated that all the participants of the experiment followed the robot’s instruction. In both experiments participants trusted the robots for different reasons. For example, some of them believed it was all staged, others that they were supposed to follow it because they accepted to participate in the experiments.

Both Salem et al. (2015) and Robinette et al. (2016)’s works showed that some participants did believe that they were acting according to the experimenter deci- sions and that their lives where not in danger. Therefore, it is still not clear from

Chapter 2 A. Rossi these results whether faulty robots are trusted by humans, and whether humans can believe that robots can look after their safety and well-being.

Research questions

Reviewing the literature regarding trust in HRI it is clear that none of the studies consider the magnitude of robot errors, nor the possibility for a robot to initiate a trust recovery process to earn back people’s trust similar to that of a human-human trust recovery in romantic, working, family or other type of relationship (Munjal Desai et al. 2013; Muir et al.1996; Robinette, W. Li, et al.2016; Maha Salem, Gabriella Lakatos, et al. 2015). In particular, several research questions have been raised that need to be investigated in order to provide robot’s behaviours and mechanisms aimed to act in case of an error and to re-gain a loss of trust. The first question to answer is RQ-1 - How do various type of robot errors affect human’s trust in a robot?, and it aims to identify how the magnitude and the timing in which robots’ errors happen affect people’s trust in a robot.

In particular, I believed that people’s trust can be affected differently depending on the magnitude of the error consequences. However, people are different from each other, in terms of personality, background, perceptions of the environments and in numerous other ways. I expect, for this reason, that they might perceive the errors made by a robot differently, hence that the same robot error may affect them differently. Therefore, it is essential to classify errors according their perceived severity of error consequences. my first research question is RQ1.1 - How do people classify errors?

Current literature focused on robots’ errors did not distinguish different magnitude of consequences (severe or limited). I expect that people’s trust might be affected differently depending on the severity level of the robots’ errors. I also believe that there is a correlation between the timing in which a robot makes an error and the loss of trust of the human in the robot. This research has been guided by the following research questions:

RQ-1.2 - Which kind of erroneous behaviours impact a human’s trust in a robot?

RQ-1.3 - Does the impact on trust change if the error happens at the beginning or end of an interaction?

RQ-1.4 - How does the combination of timing and frequency of trust have any significant impact on trust recovery process?