Gestures are a convenient modality of conveying human intent in collaborative human-robot tasks, and the pointing gesture is commonly used in pick-and-place tasks. But, it is hard to accurately detect the location pointed to with stereo cameras, and experiments in the literature tend to space out the objects of interest in order to make the task easier. We propose the use of gestures conveying spatial directions as an alternative to the pointing gesture when objects are closely packed, since inaccuracies in the detection of the spatial location pointed to can increase task completion difficulty. Using a human study, we confirmed that the gestures we propose are naturally used by humans collaborating with other humans when performing the task. Then, we develop a computer vision pipeline capable of generating a vector representing the pointing direction, and detecting specific spatial gestures from an RGB-D video stream. Using a self report survey, we show statistically significant evidence that subjects report higher satisfaction and better team performance when using spatial gestures instead of the pointing gesture to communicate with a robotic teammate. Finally, we show preliminary evidence that this trend holds true even when the accuracy of the pointing location detection is artificially inflated.

Risk is traditionally described as the expected likelihood of an undesirable outcome, such as a collision for an autonomous vehicle. Accurately predicting risk or potentially risky situations is critical for the safe operation of an autonomous vehicle. This work combines use of a controller trained to navigate around individuals in a crowd and a risk-based decision-making framework for an autonomous vehicle that integrates high-level risk-based path planning with a reinforcement learning-based low-level control. We evaluated our method using a high-fidelity simulation environment. We show our method results in zero collisions with pedestrians and predicted the least risky path, time to travel, or day to travel in approximately 72% of traversals. This work can improve safety by allowing an autonomous vehicle to one day avoid and react to risky situations.

Deep learning has achieved remarkable success in object recognition tasks through the availability of large scale datasets like ImageNet. However, deep learning systems suffer from catastrophic forgetting when learning incrementally without replaying old data. For real-world applications, robots also need to incrementally learn new objects. Further, since robots have limited human assistance available, they must learn from only a few examples. However, very few object recognition datasets and benchmarks exist to test incremental learning capability for robotic vision. Further, there is no dataset or benchmark specifically designed for incremental object learning from a few examples. To fill this gap, we present a new dataset termed F-SIOL-310 (Few-Shot Incremental Object Learning) which is specifically captured for testing few-shot incremental object learning capability for robotic vision. We also provide benchmarks and evaluations of 8 incremental learning algorithms on F-SIOL-310 for future comparisons. Our results demonstrate that the few-shot incremental object learning problem for robotic vision is far from being solved.

Robots may soon play a role in higher education by augmenting learning environments and managing interactions between instructors and learners. Little, however, is known about how the presence of robots in the learning environment will influence academic integrity. This study therefore investigates if and how college students cheat while engaged in a collaborative sorting task with a robot. We employed a 2x2 factorial design to examine the effects of cheating exposure (exposure to cheating or no exposure) and task clarity (clear or vague rules) on college student cheating behaviors while interacting with a robot. Our study finds that prior exposure to cheating on the task significantly increases the likelihood of cheating. Yet, the tendency to cheat was not impacted by the clarity of the task rules. These results suggest that normative behavior by classmates may strongly influence the decision to cheat while engaged in an instructional experience with a robot.

The capabilities and acceptance of social robots would be greatly improved by developing their ability to determine human intent. Trust, which is an important consideration in collaborations, is affected by the intent of the agents involved. Prediction of a human’s intent to deceive a robot is understudied and an important factor in calculating trust. We present a method that predicts the intent to deceive in a card game scenario that relies on facial emotion recognition. Video data collected through a human study where subjects played against three opponent types: Human, Robot, and Computer simulation, was used to validate model performance in the wild. We show that our method which uses Mini-batched K-Means not only facilitates online learning, but exceeds human performance in 17 out of 30 trials. In addition, this end-to-end pipeline removes the dependency on human annotations used in a prior method allowing for deployment on a robot.

The COVID-19 pandemic will have a profound and long-lasting impact on the entire scientific endeavor. Scientists already are adapting research programs to adapt to changes in what is prioritized—and what is possible; educators are changing the way that the next generation of researchers are trained, and flagship conferences in many fields are being cancelled, postponed, and fundamentally transformed. These broad-reaching changes are particularly impactful to human-oriented domains such as human-robot interaction (HRI). Because in-person human-subject experiments can take a year or more to conduct, the research we will see published in the field in the immediate future may appear to be “business as usual,” with accounts of laboratory studies with large numbers of in-person participants. The research currently being performed, however, is of course a different story entirely. Studies that were under way when the current crisis began will be truncated, resulting either in work that cannot be published or in work whose true impact is difficult to accurately assess. Yet HRI research performed in the coming years will be changed in fundamentally different ways; the inability to perform—or expect future performance of—in-person human subjects research, especially research involving tactile or multiparty interaction, will change both the dominant methodological techniques employed by HRI researchers and the very research questions that the field chooses to—and is able to—address. These challenges demand that HRI researchers identify precisely how the field can maintain research quality and impact while the ability to conduct human-subject studies is severely impaired for an undetermined amount of time. A natural inclination may be simply to wait the crisis out in the hope of a speedy return to normalcy; however, in this article, we argue that the community can also take this opportunity to reevaluate and refocus how research in this field is conducted and how students are mentored in ways that will yield benefits for years to come after the current crisis has ended.

This article examines how people respond to robot-administered verbal and physical punishments. Human participants were tasked with sorting colored chips under time pressure and were punished by a robot when they made mistakes, such as inaccurate sorting or sorting too slowly. Participants were either punished verbally by being told to stop sorting for a fixed time, or physically, by restraining their ability to sort with an in-house crafted robotic exoskeleton. Either a human experimenter or the robot exoskeleton administered punishments, with participant task performance and subjective perceptions of their interaction with the robot recorded. The results indicate that participants made more mistakes on the task when under the threat of robot-administered punishment. Participants also tended to comply with robot-administered punishments at a lesser rate than human-administered punishments, which suggests that humans may not afford a robot the social authority to administer punishments. This study also contributes to our understanding of compliance with a robot and whether people accept a robot’s authority to punish. The results may influence the design of robots placed in authoritative roles and promote discussion of the ethical ramifications of robot-administered punishment.

For many applications, robots will need to be incrementally trained to recognize the specific objects needed for an application. This paper presents a practical system for incrementally training a robot to recognize different object categories using only a small set of visual examples provided by a human. The paper uses a recently developed state-of-the-art method for few-shot incremental learning of objects. After learning the object classes incrementally, the robot performs a table cleaning task organizing objects into categories specified by the human. We also demonstrate the system's ability to learn arrangements of objects and predict missing or incorrectly placed objects. Experimental evaluations demonstrate that our approach achieves nearly the same performance as a system trained with all examples at one time (batch training), which constitutes a theoretical upper bound.

Human drivers continually use social information to inform their decision making. We believe that incorporating this information into autonomous vehicle decision making would improve performance and importantly safety. This paper investigates how information in the form of pedestrian density can be used to identify the path being travelled and predict the number of pedestrians that the vehicle will encounter along that path in the future. We present experiments which use camera data captured while driving to evaluate our methods for path recognition and pedestrian density prediction. Our results show that we can identify the vehicle's path using only pedestrian density at 92.4% accuracy and we can predict the number of pedestrians the vehicle will encounter with an accuracy of 70.45%. These results demonstrate that pedestrian density can serve as a source of information both perhaps to augment localization and for path risk prediction.

We present an approach for robot learning from demonstration and communication applied to simple board games like Connect Four. In such games, a visual representation of a winning condition on the board can be converted to an extensive form representation that can then support computation of a winning strategy. We present a robot that can learn simple games from responses to visual questions based on synthesized images, or to verbal questions. We illustrate how reliance on both modalities leads to more efficient learning.

The HRI community is working to develop interactive robots for a wide variety of pro-social tasks and ideals. As such we naturally focus on the positive side of HRI including how robots and humans may collaborate and the benefits of doing so. This workshop, in contrast, will focus on the dark side of HRI with the goal of identifying, understanding and guarding against the potential negative consequences of interactive robots. The primary objective of the workshop is to articulate and discuss the most pertinent ethical issues facing the HRI community and to develop a set of common community guidelines.

Robotic rescuers digging through rubble, fire-fighting drones flying over populated areas, robotic servers pouring hot coffee for you, and a nursing robot checking your vitals are all examples of current or near-future situations where humans and robots are expected to interact in a dangerous situation. Dangerous HRI is an as-yet understudied area of the field. We define dangerous HRI as situations where humans experience some amount of risk of bodily harm while interacting with robots. This interaction could take many forms, such as a bystander (e.g. when an autonomous car waits at a crossing for a pedestrian), as a recipient of robotic assistance (rescue robots), or as a teammate (like an autonomous robot working with a SWAT team). To facilitate better study of this area, the Dangerous HRI workshop brings together researchers who perform experiments with some risk of bodily harm to participants and discuss strategies for mitigating this risk while still maintaining validity of the experiment. This workshop does not aim to tackle the general problem of human safety around robots, but instead focused on guidelines for and experience from experimenters.

Recent efforts in human-robot interaction research has shed some light on the impact of human-robot interactions on human decisions during emergencies. It has been shown that presence of crowds during emergencies can influence evacuees to follow the crowd to find an exit. Research has shown that robots can be effective in guiding humans during emergencies and can reduce this `follow the crowd' behavior potentially providing life-saving benefit. These findings make robot guided evacuation methodologies an important area to explore further. In this paper we propose techniques that can be used to design effective evacuation methods. We explore the different strategies that can be employed to help evacuees find an exit sooner and avoid over-crowding to increase their chances of survival. We study two primary strategies, 1) shepherding method and 2) handoff method. Simulated experiments are performed to study the effectiveness of each strategy. The results show that shepherding method is more effective in directing people to the exit.

A number of important human-robot applications demand trust. Although a great deal of research has examined how and why people trust robots, less work has explored how robots might decide whether to trust humans. Surface cues are perceptual clues that provide hints as to a person's intent and are predictive of behavior. This paper proposes and evaluates a model for recognizing trust surface cues by a robot and predicting if a person's behavior is deceitful in the context of a trust game. The model was tested in simulation and on a physical robot that plays an interactive card game. A human study was conducted where subjects played the game against a simulation, the robot, and a human opponent. Video data was hand coded by two coders with an inter-rater reliability of 0.41 based on Levenshtein distance. It was found that the model outperformed/matched the human coders on 50% of the subjects. Overall, this paper contributes a method that may begin to allow robots to evaluate the surface cues generated by a person to determine whether or not it should trust them.

The advent and widespread adoption of wearable cameras and autonomous robots raises important issues related to privacy. The mobile cameras on these systems record and may re-transmit enormous amounts of video data that can then be used to identify, track, and characterize the behavior of the general populous. This paper presents a preliminary computational architecture designed to preserve specific types of privacy over a video stream by identifying categories of individuals, places, and things that require higher than normal privacy protection. This paper describes the architecture as a whole as well as preliminary results testing aspects of the system. Our intention is to implement and test the system on ground robots and small UAVs and demonstrate that the system can provide selective low-level masking or deletion of data requiring higher privacy protection.

This article presents a conceptual framework for human-robot trust which uses computational representations inspired by game theory to represent a definition of trust, derived from social psychology. This conceptual framework generates several testable hypotheses related to human-robot trust. This article examines these hypotheses and a series of experiments we have conducted which both provide support for and also conflict with our framework for trust. We also discuss the methodological challenges associated with investigating trust. The article concludes with a description of the important areas for future research on the topic of human-robot trust.

Robots have the potential to save lives in high-risk situations, such as emergency evacuations. To realize this potential, we must understand how factors such as the robot's performance, the riskiness of the situation, and the evacuee's motivation influence his or her decision to follow a robot. In this paper, we developed a set of experiments that tasked individuals with navigating a virtual maze using different methods to simulate an evacuation. Participants chose whether or not to use the robot for guidance in each of two separate navigation rounds. The robot performed poorly in two of the three conditions. The participant's decision to use the robot and self-reported trust in the robot served as dependent measures. A 53% drop in self-reported trust was found when the robot performs poorly. Self-reports of trust were strongly correlated with the decision to use the robot for guidance (φ(90) = +0.745). We conclude that a mistake made by a robot will cause a person to have a significantly lower level of trust in it in later interactions.

Robots have the potential to save lives in emergency scenarios, but could have an equally disastrous effect if participants overtrust them. To explore this concept, we performed an experiment where a participant interacts with a robot in a non-emergency task to experience its behavior and then chooses whether to follow the robot's instructions in an emergency or not. Artificial smoke and fire alarms were used to add a sense of urgency. To our surprise, all 26 participants followed the robot in the emergency, despite half observing the same robot perform poorly in a navigation guidance task just minutes before. We performed additional exploratory studies investigating different failure modes. Even when the robot pointed to a dark room with no discernible exit the majority of people did not choose to safely exit the way they entered.

Most Human-Robot Interaction (HRI) experiments are costly and time consuming because they involve deploying a physical robot in a physical space. Experiments using virtual environments can be easier and less expensive, but it is difficult to ensure that the results will be valid in the physical domain. To begin to answer this concern, we have performed an evaluation comparing participants' understanding of robotic guidance instructions using robots that were virtually, remotely, or physically present for the experiment. All but one set of experimental conditions gave similar results across the three presence levels. Further, we find that qualitative responses about the robots were largely the same regardless of presence level.

simultaneously licensed under a Creative Commons Attribution License that allows others to share the work with an acknowledgement of the work's authorship and initial publication in this journal.

Motivated by the desire to mitigate human casualties in emergency situations, this paper explores various guidance modalities provided by a robotic platform for instructing humans to safely evacuate during an emergency. We focus on physical modifications of the robot, which enables visual guidance instructions, since auditory guidance instructions pose potential problems in a noisy emergency environment. Robotic platforms can convey visual guidance instructions through motion, static signs, dynamic signs, and gestures using single or multiple arms. In this paper, we discuss the different guidance modalities instantiated by different physical platform constructs and assess the abilities of the platforms to convey information related to evacuation. Human-robot interaction studies with 192 participants show that participants were able to understand the information conveyed by the various robotic constructs in 75.8% of cases when using dynamic signs with multi-arm gestures, as opposed to 18.0% when using static signs for visual guidance. Of interest to note is that dynamic signs had equivalent performance to single-arm gestures overall but drastically different performances at the two distance levels tested. Based on these studies, we conclude that dynamic signs are important for information conveyance when the robot is in close proximity to the human but multi-arm gestures are necessary when information must be conveyed across a greater distance.

Psychologists note that humans regularly use categories to simplify and speed up the process of person perception [1]. The influence of categorical thinking on interpersonal expectations is commonly referred to as a stereotype. The ability to bootstrap the process of learning about a newly encountered, unknown person is critical for robots interacting in complex and dynamic social situations. This article contributes a novel cluster-based algorithm that allows a robot to create generalized models of its interactive partner. These generalized models, or stereotypes, act as a source of information for predicting the human's behavior and preferences. We show, in simulation and using real robots, that these stereotyped models of the partner can be used to bootstrap the robot's learning about the partner in spite of significant error. The results of this work have potential implications for social robotics, autonomous agents, and possibly psychology.

As humans are being progressively pushed further downstream in the decision-making process of autonomous systems, the need arises to ensure that moral standards, however defined, are adhered to by these robotic artifacts. While meaningful inroads have been made in this area regarding the use of ethical lethal military robots, including work by our laboratory, these needs transcend the warfighting domain and are pervasive, extending to eldercare, robot nannies, and other forms of service and entertainment robotic platforms. This paper presents an overview of the spectrum and specter of ethical issues raised by the advent of these systems, and various technical results obtained to date by our research group, geared towards managing ethical behavior in autonomous robots in relation to humanity. This includes: 1) the use of an ethical governor capable of restricting robotic behavior to predefined social norms; 2) an ethical adaptor which draws upon the moral emotions to allow a system to constructively and proactively modify its behavior based on the consequences of its actions; 3) the development of models of robotic trust in humans and its dual, deception, drawing on psychological models of interdependence theory; and 4) concluding with an approach towards the maintenance of dignity in human-robot relationships.

This article presents an investigation into the theoretical and computational aspects of trust as applied to robots. It begins with an in-depth review of the trust literature in search of a definition for trust suitable for implementation on a robot. Next we apply the definition to our interdependence framework for social action selection and develop an algorithm for determining if an interaction demands trust on the part of the robot. Finally, we apply our algorithm to several canonical social situations and review the resulting indications of whether or not the situation demands trust.

A 3D optical flow tracking system was used to track participants as they watched a series of boring videos. The video stream of the participants was rated for boredom events. Ratings and head position data were combined to predict boredom events.

This paper explores the use of an outcome matrix as a computational representation of social interaction suitable for implementation on a robot. An outcome matrix expresses the reward afforded to each interacting individual with respect to pairs of potential behaviors. We detail the use of the outcome matrix as a representation of interaction in social psychology and game theory, discuss the need for modeling the robot's interactive partner, and contribute an algorithm for creating outcome matrices from perceptual information. Experimental results explore the use of the algorithm with different types of partners and in different environments.

As the capabilities, range of missions, and the size of robot teams increase, the ability for a human operator to account for all the factors in these complex scenarios can become exceedingly difficult. Our previous research has studied the use of case-based reasoning (CBR) tools to assist a user in the generation of multi-robot missions. These tools, however, typically assume that the robots available for the mission are of the same type (i.e., homogeneous). We loosen this assumption through the integration of contract-net protocol (CNP) based task allocation coupled with a CBR-based mission specification wizard. Two alternative designs are explored for combining case-based mission specification and CNP-based team allocation as well as the tradeoffs that result from the selection of one of these approaches over the other.

This paper presents a theoretical framework for computationally representing social situations in a robot. This work is based on interdependence theory, a social psychological theory of interaction and social situation analysis. We use interdependence theory to garner information about the social situations involving a human and a robot. We also quantify the gain in outcome resulting from situation analysis. Experiments demonstrate the utility of social situation information and of our situation-based framework as a method for guiding robot interaction. We conclude that this framework offers a principled, general approach for studying interactive robotics problems

This paper presents a method for multi-robot communication sensitive reconnaissance. This approach utilizes collections of precompiled vector fields in parallel to coordinate a team of robots in a manner that is responsive to communication failures. Collections of vector fields are organized at the task level for reusability and generality. Different team sizes, scenarios, and task management strategies are investigated. Results indicate an acceptable reduction in communication attenuation when compared to other related methods of navigation. Online management of tasks and potential scalability are discussed.

Autonomous teams of robots operating in a dynamic, adversarial environment stand to benefit from using all available resources. But how can knowledge be used to construct a plan that does not interfere with the robots ability to react to its environment? In this research we distill abstract representation into a plan usable by a reactive behavior-based architecture. This plan is then exploited to enhance the performance of a team of robots tasked with maintaining communications while performing reconnaissance. Utilizing multiple plans in serial and in parallel is shown via simulation to be a promising method for increasing mission performance. We conclude that the utility of these internalized plans warrants further investigation as a method for imbuing reactive agents with a priori knowledge.