Despite advancements in socially aware navigation, robots still often behave inappropriately in social environments. To ensure successful application, robots must detect the human perceived appropriateness of their navigation behaviors. This paper presents a novel dataset covering a complete range of perceived appropriateness and uniquely incorporates human emotion and attention to facilitate the detection of perceived appropriateness of robot social navigation in pathways (PARSNiP). It is created based on a series of human-robot interaction experiments with 30 participants and a mobile robot. Several typical machine learning models are utilized to evaluate the dataset and analyze the contributions of different features in detecting perceived appropriateness. The results indicate that incorporating emotional and attentional features can significantly improve the accuracy of perceived appropriateness detection. There was an increase from 63% to 68% using algorithm-predicted emotional and attentional features, and a further increase to 79% with the emotion and attention data reported by the participants. With the dataset, researchers could train machine learning models to enable robots to detect perceived appropriateness accurately, fostering adaptations that improve their responsiveness and accuracy in social interactions. The dataset is available for download at https://github.com/duibcuiegiosahxois/PARSNiP.git, and videos will be shared upon request by contacting Y.Zhou-13@tudelft.nl.

Robots are increasingly navigating our living environments and must navigate socially to be accepted. While existing socially aware navigation (SAN) approaches enable robots to interpret and communicate social information to navigate efficiently, safely, and in a socially acceptable manner, they often overlook potential conflicts and errors in real-world human-robot interactions. This thesis contributes to SAN by investigating how robots can adapt their inappropriate navigation behavior based on human feedback (perceived appropriateness), leading to smoother and less error-prone human-robot interactions....

In social environment navigation, robots inevitably exhibit behaviors that are perceived as inappropriate by humans. Current robots lack the ability to adapt to such human perceptions, leading to repeated inappropriate behaviors. This study employs a mixed-methods approach to explore human-preferred robot adaptations, combining qualitative data from a series of human-robot interactions and a semi-structured interview, and quantitative data from an online survey. 12 participants were recruited to interact with a mobile robot in an indoor setting, reporting 139 instances of inappropriate robot behaviors. The subsequent semi-structured interviews regarding these instances yielded 9 types of inappropriate behaviors and 10 major types of human-preferred robot adaptations, ranging from general ones, such as stopping the motion, to more specific ones, like moving away and then stopping. Additionally, 12 human-preferred adaptations were selected from the interview data and presented to the same participants through an online survey to evaluate their effectiveness in addressing the inappropriate behaviors previously identified. The results reveal the human preference for the robot to move to the side and then stop in most scenarios, which might serve as a general adaptation for addressing inappropriate robot navigation behaviors.

Robots navigating in social environments inevitably exhibit behavior perceived as inappropriate by people, which they will repeat unless they are aware of them; hindering their social acceptance. This highlights the importance of robots detecting and adapting to the perceived appropriateness of their behavior, in line with what we found in a systematic literature review. Therefore, we have conducted experiments (both outdoor and indoor) to understand the perceived appropriateness of robot social navigation behavior, based on which we collected a dataset and developed a machine learning model for detecting such perceived appropriateness. To investigate the usefulness of such information and inspire robot adaptive navigation behavior design, we will further conduct aWoZ study to understand how trained human operators adapt robot behavior to people's feedback. In all, this work will enable robots to better remediate their inappropriate behavior, thus improving their social acceptance.

When mobile urban robots will share the sidewalk with people, the resulting interactions can cause unexpected undesirable outcomes to emerge – from people running away scared to people deliberately teasing and harassing such systems. How can we design such AI systems to aptly handle the unexpected? Directly anticipating and/or detecting these kinds of situations will inherently be unreliable; they are unexpected, after all. And yet, there exists a very clear signal for social slip-ups: the emotional response of people. We thus argue that such systems need to be imbued with a capacity to interpret the socio-emotional reactions to their own behavior.