Beyond R2D2 - The design of nonverbal interaction behavior optimized for robot-specific morphologies

It is likely that in the near future we will meet more and more robots that will perform tasks in social environments, such as shopping malls, airports or museums. However, design guidelines that inform the design of effective nonverbal behavior for robots are scarce. This is surprising since the behavior of robots in social environments can greatly affect the interaction between people and robots. Previous work has shown that people have the tendency to project motivations, intentions and emotions on non-living objects, such as robots, when these objects start to move. Moreover, it has been shown that people tend to react socially to computers, televisions and new media. Similarly, people will react socially to robots. Currently, in the field of human-robot interaction the focus is on creating robots that look like humans and that can use humanlike behavior in interaction. However, such robots are not suitable for all tasks, and humanlike robots are complex, vulnerable and expensive. Moreover, people do not always prefer humanlike robots over robots with other appearances. This indicates that there are good reasons to develop low-anthropomorphic robots; robots that do not resemble people in appearance. A challenge in designing nonverbal behavior for low- anthropomorphic robots is that these robots often lack the abilities to imitate humanlike behavior correctly. These robots could lack the specific modalities to perform the behavior, for example, they may not have arms and fingers to point with. When a robot does have the right modalities to perform a specific behavior, these modalities might not have similar degrees of freedom to imitate recognizable nonverbal behavior. Yet, contrary to people, low-anthropomorphic robots may use screens, projections, light cues or specific movements to communicate their intentions and motivations. To optimize the use of robot-specific modalities and morphology, we developed robot-optimized behavior. This was done in the context of developing guide robots for tourist sites. To understand the effects of both imitated humanlike behavior and robot-optimized behavior for low-anthropomorphic robots on peoples’ perception, controlled lab studies and in-the-wild studies have been performed. First, we analyzed the effect and effectiveness of imitated humanlike guide behavior on low-anthropomorphic robots. These studies showed that humanlike behavior is preferred over random behavior, but that it is also more distracting. This means that humanlike behavior may not the best solution to design behavior for low-anthropomorphic robots. An important question now was what would be a promising alternative to imitating human behavior. In follow-up studies we compared the effect and effectiveness of imitated humanlike behavior for the robot to robot- optimized behavior. From these studies we learned that robot-optimized behavior is a good alternative for low-anthropomorphic robots instead of imitated humanlike behavior. To effectively perform the studies mentioned above, we developed and introduced DREAM, Data Reduction Event Analysis Method. This method allowed us to analyze video data of in-the-wild human-robot interaction. Such a method did not yet exist, and is essential in order to gain insight in effective behaviors for robots. Because many more aspects than only the behavior of the robot and the person interacting with the robot play a role in the final user experience, such as context and other people in that context, it is necessary to add real world field studies to studies in controlled (lab) environments. Therefore, in this thesis, DREAM, is introduced to analyze the rich and unstructured data of in the wild human-robot interactions in a fast and reliable manner. This method turned out to be effective to use for the guide context human-robot interaction data. The work presented in this thesis is a first step towards understanding 1) the effect of nonverbal behavior for low-anthropomorphic robots, 2) which design approaches can be effectively used to design this behavior and 3) how nonverbal robot behavior can be studied and evaluated in the wild. The tangible outcomes of this thesis: 1) the robot-optimized approach to design behavior for low-anthropomorphic robots, and 2) DREAM to evaluate human-robot interaction in the wild. These results can well serve as a starting point to further develop more diverse and effective design approaches in human-robot interaction.