We conducted a user study for which we purposefully programmed faulty behavior into a robot's routine. It was our aim to explore if participants rate the faulty robot different from an error-free robot and which reactions people show in interaction with a faulty robot. The study was based on our previous research on robot errors where we detected typical error situations and the resulting social signals of our participants during social human-robot interaction. In contrast to our previous work, where we studied video material in which robot errors occurred unintentionally, in the herein reported user study, we purposefully elicited robot errors to further explore the human interaction partners' social signals following a robot error. Our participants interacted with a human-like NAO, and the robot either performed faulty or free from error. First, the robot asked the participants a set of predefined questions and then it asked them to complete a couple of LEGO building tasks. After the interaction, we asked the participants to rate the robot's anthropomorphism, likability, and perceived intelligence. We also interviewed the participants on their opinion about the interaction. Additionally, we video-coded the social signals participants showed during their interaction with the robot as well as the answers they provided the robot with. Our results show that participants liked the faulty robot significantly better than the robot that interacted flawlessly. We did not find significant differences in people's rating of the robot's anthropomorphism and perceived intelligence. The qualitative data confirmed the questionnaire results in showing that although the participants recognized the robot's mistakes, they did not necessarily reject the erroneous robot. The annotations of the video data further showed that gaze shifts (e.g., from and object to the robot or vice versa) and laughter are typical reactions to unexpected robot behavior. In contrast to existing research, we assess dimensions of user experience that have not been considered so far and we analyze the reactions users express when a robot makes a mistake. Our results show that decoding a human's social signals can help the robot understand that there is an error and subsequently react accordingly.

Human–robot interactions are often affected by error situations that are caused by either the robot or the human. Therefore, robots would profit from the ability to recognize when error situations occur. We investigated the verbal and non-verbal social signals that humans show when error situations occur in human–robot interaction experiments. For that, we analyzed 201 videos of five human–robot interaction user studies with varying tasks from four independent projects. The analysis shows that there are two types of error situations: social norm violations and technical failures. Social norm violations are situations in which the robot does not adhere to the underlying social script of the interaction. Technical failures are caused by technical shortcomings of the robot. The results of the video analysis show that the study participants use many head movements and very few gestures, but they often smile, when in an error situation with the robot. Another result is that the participants sometimes stop moving at the beginning of error situations. We also found that the participants talked more in the case of social norm violations and less during technical failures. Finally, the participants use fewer non-verbal social signals (for example smiling, nodding, and head shaking), when they are interacting with the robot alone and no experimenter or other human is present. The results suggest that participants do not see the robot as a social interaction partner with comparable communication skills. Our findings have implications for builders and evaluators of human–robot interaction systems. The builders need to consider including modules for recognition and classification of head movements to the robot input channels. The evaluators need to make sure that the presence of an experimenter does not skew the results of their user studies.

This paper describes a context-dependent social gaze-control system implemented as part of a humanoid social robot. The system enables the robot to direct its gaze at multiple humans who are interacting with each other and with the robot. The attention mechanism of the gaze-control system is based on features that have been proven to guide human attention: nonverbal and verbal cues, proxemics, the visual field of view, and the habituation effect. Our gaze-control system uses Kinect skeleton tracking together with speech recognition and SHORE-based facial expression recognition to implement the same features. As part of a pilot evaluation, we collected the gaze behavior of 11 participants in an eye-tracking study. We showed participants videos of two-person interactions and tracked their gaze behavior. A comparison of the human gaze behavior with the behavior of our gaze-control system running on the same videos shows that it replicated human gaze behavior 89% of the time.

A robot agent designed to engage in real-world human–robot joint action must be able to understand the social states of the human users it interacts with in order to behave appropriately. In particular, in a dynamic public space, a crucial task for the robot is to determine the needs and intentions of all of the people in the scene, so that it only interacts with people who intend to interact with it. We address the task of estimating the engagement state of customers for a robot bartender based on the data from audiovisual sensors. We begin with an offline experiment using hidden Markov models, confirming that the sensor data contains the information necessary to estimate user state. We then present two strategies for online state estimation: a rule-based classifier based on observed human behaviour in real bars, and a set of supervised classifiers trained on a labelled corpus. These strategies are compared in offline cross-validation, in an online user study, and through validation against a separate test corpus. These studies show that while the trained classifiers are best in a cross-validation setting, the rule-based classifier performs best with novel data; however, all classifiers also change their estimate too frequently for practical use. To address this issue, we present a final classifier based on Conditional Random Fields: this model has comparable performance on the test data, with increased stability. In summary, though, the rule-based classifier shows competitive performance with the trained classifiers, suggesting that for this task, such a simple model could actually be a preferred option, providing useful online performance while avoiding the implementation and data-scarcity issues involved in using machine learning for this task.

In this paper, the characteristics and performance of three open-source simulators for robotics, V-REP, Gazebo and ARGoS, are thoroughly analysed and compared. While they all allow for programming in C++, they also represent clear alternatives when it comes to the trade-off between complexity and performance. Attention is given to their built-in features, robot libraries, programming methods and the usability of their user interfaces. Benchmark test results are reported in order to identify how well the simulators can cope with environments of varying complexity. The richness of features of V-REP and the strong performance of Gazebo and ARGoS in complex scenes are highlighted. Various usability issues of Gazebo are also noted.

Human activity recognition is crucial for intuitive cooperation between humans and robots. We present an approach for activity recognition for applications in the context of human-robot interaction in industrial settings. The approach is based on spatial and temporal features derived from skeletal data of human workers performing assembly tasks. These features were used to train a machine learning framework, which classifies discrete time frames with Random Forests and subsequently models temporal dependencies between the resulting states with a Hidden Markov Model. We considered the following three groups of activities: Movement, Gestures, and Object handling. A dataset has been collected which is comprised of 24 recordings of several human workers performing such activities in a human-robot interaction environment, as typically seen at small and medium-sized enterprises. The evaluation shows that the approach achieves a recognition accuracy of up to 88% for some activities and an average accuracy of 73%.