between people and robots also applies in situations in which gestural action is the main mode of interaction.
8.5.1 Effects on Perceptual Metrics
Participants found the robot’s actions easier to correct, and they found the robot sig-nificantly more compliant and predictable in the repair condition, in comparison to the condition without repair. The finding that the robot is rated as more compliant and that its behavior is rated as easier to correct are in themselves not very novel. However, for the vast majority of participants, the robot was able to show off this ability only once, under very specific circumstances, and with a quite substantial delay between repair initiation and response, due to the latency of the robot planner. Even under these tight restrictions, this one action has significant consequences. It indicates that a robot that is able to respond to repair initiations under less restrictive circumstances is likely to affect perceptions of intelligence to an even greater extent.
However, from an HRI perspective it is somewhat surprising that participants rate the robot in the repair condition as more predictable. This result indicates that responses to repair initiations not only contribute to the robot’s compliance, but also its legibility. While there are studies that have found a relation between adaptability and legibility (Dehais, Sisbot, Alami, & Causse, 2011;Moon et al., 2014) in human-robot handovers, no studies have to date reported a relation between legibility and responses to repair initiations.
However, from a conversation analytical perspective, the result makes a lot of sense. In social exchanges between people, the ability to adjust and ratify the common ground between people is taken for granted. Thus, the response to repair initiation the robot makes in therepair condition indicates the robot’s awareness toward a certain aspect of a participant’s behavior, which works as a very concrete clue as to how to instruct the robot and possibly also how not to do so. This hypothesis is also corroborated by behavioral results, which found that the response to the repair initiation provides participants with a better understanding of how to interact with the robot.
Despite the communicative breakdown created by the robot error, participants generally consider the collaboration as a success, regardless of condition, the difference is although participants in therepair condition rate the collaboration as more successful (marginally significant). A possible explanation for this can probably be found in the fact that the error committed by the robot was not fatal. Participants were able to continue the interaction even when the robot did not respond to their repair initiations. Several previous studies posit that one recovery strategy that all robots should have at their disposal is to ignore interactional trouble (Lenz et al., 2012;Opfermann & Pitsch, 2017).
8.5.2 Behavioral Effects
Analyses of participants behavior while interacting with the robot show effects both related to how participants instruct the robot and to how participants coordinate the handover of objects with the robot. The current study shows that the robot’s response to repair initiation displays to participants which aspects of their conduct the robot is aware of and
8.5 Discussion 131 responds to.
The study showed that participants in the repair condition indeed changed their instructional behavior after the robot had displayed an awareness toward their repair attempt. Prior to the error, participants in both conditions attempt to control the robot using voice, 20% of the participants in the repair condition, compared to 26.2% for participants in the no-repair condition. This distribution is comparable to other work on the same robotic platform in a similar scenario (Jensen et al., 2015), which shows that 20% of participants attempt to direct the robot using verbal commands. However, for participants in the repair condition this percentage drops drastically (see Figure 8.7). Participants in the repair condition deal with an objectively more compliant robot, but more importantly the way in which the robot responds to repair indicates to participants how they should interact with it. The robot is responsive to repair and thus adjusts the common ground between them.
That is, participants become more aware of what aspects of their communication with the robot it is aware of.
The exploratory analysis revealed two interaction formats when coordinating the handover of objects. In one format the handovers become more fluent over time, which is evidenced by a linear decrease in the time it takes to effect the handovers. However, for about one third of participants a second interaction format was revealed. Here, handovers also become more fluent over time, but in contrast to the first interaction format, the second handover is less fluent than than the first. Further analysis showed that participants in the no-repair condition were significantly more likely to follow the non-linear interaction format than participants in the repair condition. The results show that human-robot collaborations do not simply become more fluent over time, as previous work would suggest; instead, people’s expectations that the robot will build on previous interactions results in longer response times and hence less fluent interactions.
People generally bring their experience with interacting with other people to bear in interactions with non-humans, such as robots. Therefore it is logical to assume that after participants have instructed the robot how to do the handover, they expect that it would able to do this autonomously the second time. When that does not happen, interaction trouble surfaces, evidenced by long stretches of non-action. However, at the time of the second handover, participants in the repair condition will already have uncovered some of the limitations of the robot, for example that the robot only responds to gestural instructions, as discussed above. Thus, the common ground between participant and robot for participants in the repair condition are adjusted, so that the participants’ partner models of the robot reflect its abilities more accurately. Participants in the on-repair condition do not have the same possibilities of adjusting the common ground and are thus more likely to assume that the robot are able to process multiple modalities (as evidenced inFigure 8.7) and learns from instruction (as evidenced inFigure 8.10).
In summary, the studies show that implementing just one opportunity for repair can significantly affect perception and behavior. Specifically, it was shown that the response to repair serves to update participants’ partner model of the robot and subsequently changes how they interact with the robot, which methods they use, and how they perceive the
robot.
9. Discussion
The overall aim with the six empirical studies in this thesis was to explore how robots’
displays of awareness of participants, their behavior, and the context in which the interaction takes place affect interaction and how people perceive robots. Correspondingly, I have studied displays of awareness of the perceptual basis (Chapters 4 and 5), displays of awareness of the actional basis, in particular of proactivity (Chapter 7), contingency (Chapters 3and8), incrementality (Chapters 5and 6), and displays of awareness of the discourse record (Chapter 4). The underlying assumption in all of these studies is that these displays work as signals for what the robot considers common ground. Four of the five indicators were shown to adjust understandings of common ground displayed by participants.