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.
happened in previous interactions. The results fromChapter 3add to the body of work that shows that contingent gaze signals an awareness of the discourse record.
In another study, also investigating contingency,Chapter 8, contingency is implemented as a repair mechanism. More specifically, a robot is able to recognize non-verbal repair initiations and respond to them contingently. The study shows that the robot’s response to repair initiation, displays to participants which aspects of their conduct the robot is aware of and responds to. The 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, whereas in the other format the second handover takes longer to complete than the first one. Participants in theno-repair condition were significantly more likely to follow the non-linear interaction format than participants in therepair condition. As I also argue in Chapter 8 that the response to repair initiations makes it clear to participants how they should (and should not) instruct the robot. Thus, this contingent repair grounds the understanding of what the robot is able to understand and what it is able to do.
Thus, for both implementations of common ground, contingency grounds participants’
understandings of the robot’s ability, which had direct consequences for how people interacted with the robots. Contingent gaze leads participants to make assumptions about the robot’s ability to read, understand and recall from earlier on in the interaction.
Contingent repair leads participants to reevaluate some of the assumptions they had made of the robot’s ability to understand their instruction. It can therefore be said that as contingency signals an awareness of some aspect of an interaction, this signal becomes a cue through which participants understand what the robot understands (or displays to understand).
9.1.2 Incrementality
Incrementality was implemented in two different studies. Incremental feedback had been hypothesized to signal situatedness; it signals to participants that the robot understands and responds to moment-to-moment changes in participants’ behaviors. The first study on incrementality (Chapter 5) found that incrementality positively affects intelligence and trustworthiness. Other studies have also found relations between markers of competence, such as intelligence, and incremental feedback (Baumann & Lindner, 2015; Skantze &
Hjalmarsson, 2013).
The second study on incrementality (Chapter 6) found only very few differences between incremental and non-incremental feedback. Surprisingly, the robot in the incremental condition was rated as significantly less credible than the robot in the non-incremental condition. This stands in contrast to the results fromChapter 5that found that participants found the incremental robot more trustworthy than the non-incremental robot. This discrepancy is, however, not surprising. Previous works have produced results that seem to contradict each other. For example, incrementality has been reported to positively affect the extent to which participants think a robot is natural (Buschmeier et al., 2012), but other works show the opposite (Chromik et al., 2017). Likewise, one study reports that
9.1 Indicators for Common Ground 135 robots utilizing incremental speech are perceived as more enjoyable (Tsai et al., 2018), while another study reports that a robot utilizing incremental speech is perceived as less likable. For these studies, incremental speech may have been implemented in different ways, which may have had an impact participants’ perceptions. However, for the studies in Chapter 5 and Chapter 6, incrementality is implemented in very similar ways, even though the tasks and robots are different. Thus, elements in the makeup of the task could be responsible for the discrepancy. Another explanation could be that because of the difference in task, people perceive the feedback differently. The data at hand does not enable further investigations of this discrepancy. In other words, more (empirical) work needs to be done in order to properly understand the relationship between incremental feedback and perceptive effects, such as credibility and trustworthiness.
The second study on incrementality (Chapter 6) found significant differences in participant behavior between the experimental conditions. Specifically, participants in theincremental condition solved their tasks faster than participants in the non-incremental condition.
This result is in line with previous works on incremental speech, which have found that incrementality increases performance (Kennington et al., 2014;Skantze & Hjalmarsson, 2010; 2013). For participants in the non-incremental condition, the robot displays no awareness of the participants’ whereabouts or conduct. It merely states where an object can be found. As a result, there is much less common ground between participant and robot and what common ground there is, is not continually updated as is the case for participants in the incremental condition. Thus, adding incremental feedback to a robot’s communication design increases the perceived common ground between robot and participants. Analyses of interactions between performative and perceptive metrics revealed an interesting relationship. In particular, they show that in the non-incremental condition, participants’ ratings of the robot are generally unaffected by how long it takes them to find the right objects. In other words, difficulties in finding the objects are not reflected negatively on the robot. This is different for participants in the incremental condition. For those participants difficulties in locating the object are reflected by negative ratings of the degree to which participants thought the robot took them into account, the degree to which they thought the robot responded to their actions, and participants’ levels of discomfort.
Interactions between perceptual and behavioral data inChapter 6 suggest that incremental feedback may increase performance, as previous work also suggests (Chromik et al., 2017;
Ghigi et al., 2014;Kennington et al., 2014;Skantze & Hjalmarsson, 2010;2013), but the increased performance also comes at a price. In other words, participants who interact with a robot that displays an ability to process input and output incrementality hold the robot to a higher standard than participants who interact with a robot without these abilities.
Specifically, participants who interacted with the robot endowed with incremental feedback and who for some reason encountered trouble in finding the objects perceived the robot as less competent, more discomforting, felt that to a lesser degree that it responded to their actions and took them into account than participants who either did not encounter any problems or participants who interacted with the robot in the non-incremental condition.
This means that these participants hold the robot accountable for their performance. Thus,
incrementality signals an awareness to context, which affects the assumptions participants make about the robot. Incrementality grounds participants’ understandings of the robot’s displays of awareness. This leads to a more smooth interaction when assumptions are met, but can also affect the perception of the robot negatively when problems occur.
9.1.3 Discourse Record
Awareness of the discourse record is implemented in a single study in Chapter 4. The study showed that just a single reference to the discourse record affected the degree to which participants found the robot aware, social and interactive. This reference signals to participants that the discourse record is part of the common ground. Much of previous work encode statistical information (such as game scores) in memory systems for robots (Ahmad et al., 2017;Kipp & Kummert, 2016;Leite et al., 2014). While this work is relevant and important, robots that are designed to engage people in social interaction will also need the ability to be able to encode and interpret talk. Another way in which the results from Chapter 4differ from previous work is that the robot did not only make a reference to a previous utterance, but recontexualized it for the current utterance. Specifically, the robot used a response to a previous question to be used in a new question where participants are asked to evaluate their experience. While this form of memory access is likely to be more difficult to implement in an autonomous agent than, for example, game statistics, the results fromChapter 4indicate that the difficulties might be worth the work. The results also resonate with a hypotheses put forth byChristian (2011) who posits that in order for computers (and thus robots) to become more ‘human’, they need to display that they can make use of and recontextualize past experiences.
9.1.4 The Perceptual Basis
The perceptual basis as an indicator for common ground was implemented in two experi-ments, presented inChapter 4 and Chapter 5. Results show that displays of awareness influence perceived traits such as likability and authority, but also influences compliance.
Specifically, the displays of awareness of to what happens in the immediate environment support the robot in its other tasks, for example getting participants to drink more water.
The perceptual basis is largely unexplored in HRI. Much work on situation awareness in HRI that focuses on a robot’s situation awareness, rather than of a controller’s (see Chapter 1), investigates participants’ actions (e.g. Baxter et al. (2014), Ishii et al. (2013)) rather than the environment participants are in. While this work is relevant and very important, I argue that looking in the immediate space in which interaction takes place may offer equally valuable cues to what an interaction partner might be doing or engaged in. Especially, the study presented in Chapter 5 showed that displays of awareness of the perceptual basis can affect perception and interaction positively. As an indicator for common ground in HRI, the displays of awareness to the perceptual basis signal a joint understanding of physical environment in which interaction takes place.