We compare three methods of familiarizing a human with an artificial intelligence (AI) teammate ("agent") prior to operation in a collaborative, fast-paced intelligence, surveillance, and reconnaissance (ISR) environment. In a between-subjects user study (n=60), participants either read documentation about the agent, trained alongside the agent prior to the mission, or were given no familiarization. Results showed that the most valuable information about the agent included details of its decision-making algorithms and its relative strengths and weaknesses compared to the human. This information allowed the familiarization groups to form sophisticated team strategies more quickly than the control group. Documentation-based familiarization led to the fastest adoption of these strategies, but also biased participants towards risk-averse behavior that prevented high scores. Participants familiarized through direct interaction were able to infer much of the same information through observation, and were more willing to take risks and experiment with different control modes, but reported weaker understanding of the agent’s internal processes. Significant differences were seen between individual participants’ risk tolerance and methods of AI interaction, which should be considered when designing human-AI control interfaces. Based on our findings, we recommend a human-AI team familiarization method that combines AI documentation, structured in-situ training, and exploratory interaction.

We investigate the real-time estimation of human situation awareness using observations from a robot teammate with limited visibility. In human factors and human-autonomy teaming, it is recognized that individuals navigate their environments using an internal mental simulation, or mental model. The mental model informs cognitive processes including situation awareness, contextual reasoning, and task planning. In teaming domains, the mental model includes a team model of each teammate’s beliefs and capabilities, enabling fluent teamwork without the need for explicit communication. However, little work has applied team models to human-robot teaming. We compare the performance of two current methods at estimating user situation awareness over varying visibility conditions. Our results indicate that the methods are largely resilient to low-visibility conditions in our domain, however opportunities exist to improve their overall performance.

This paper presents an incremental replanning algorithm, dubbed LTL-D*, for temporal-logic-based task planning in a dynamically changing environment. Unexpected changes in the environment may lead to failures in satisfying a task specification in the form of a Linear Temporal Logic (LTL). In this study, the considered failures are categorized into two classes: (i) the desired LTL specification can be satisfied via replanning, and (ii) the desired LTL specification is infeasible to meet strictly and can only be satisfied in a "relaxed" fashion. To address these failures, the proposed algorithm finds an optimal replanning solution that minimally violates desired task specifications. In particular, our approach leverages the D* Lite algorithm and employs a distance metric within the synthesized automaton to quantify the degree of the task violation and then replan incrementally. This ensures plan optimality and reduces planning time, especially when frequent replanning is required. Our approach is implemented in a robot navigation simulation to demonstrate a significant improvement in the computational efficiency for replanning by two orders of magnitude.

Trust is crucial for technological acceptance, continued usage, and teamwork. However, human-robot trust, and human-machine trust more generally, suffer from terminological disagreement and construct proliferation. By comparing, mapping, and analyzing well-constructed trust survey instruments, this work uncovers a consensus structure of trust in human–machine interaction. To do so, we identify the most frequently cited and best-validated human-machine and human-robot trust questionnaires as well as the best-established factors that form the dimensions and antecedents of such trust. To reduce both confusion and construct proliferation, we provide a detailed mapping of terminology between questionnaires. Furthermore, we perform a meta-analysis of the regression models which emerged from the experiments that employed multi-factorial survey instruments. Based on this meta-analysis, we provide the most complete, experimentally validated model of human-machine and human-robot trust to date. This convergent model establishes an integrated framework for future research. It determines the current boundaries of trust measurement and where further investigation and validation are necessary. We close by discussing how to choose an appropriate trust survey instrument and how to design for trust. By identifying the internal workings of trust, a more complete basis for measuring trust is developed that is widely applicable.

Human-robot teams can leverage a human’s expertise and a robot’s computational power to meaningfully improve mission outcomes. In command and control domains, the robot teammate can also act as a decision-support system to advise human users. However, decision-support systems are susceptible to human factors issues including miscalibrated trust and degraded team performance. Recent work has mitigated these issues by using cognitive forcing functions to structure shared decision-making systems and place users as proactive on-the-loop actors. We bring this approach to a human-robot teaming domain, and investigate how Type I and Type II errors in the robot’s recommendation affects team performance and user rational trust. We present the architecture of our decision-making process and a Mars rover landing experiment domain. Results from a comprehensive user study demonstrate that the error type of the robot’s recommendation forms a trade-off between team performance and rational trust.

Understanding human mental models of AI are critical for designing human-centered AI. Examining mental models provides depth in understanding of how users want to interact with AI, when users may need additional explanation of the system, and what knowledge is shared between the user and the AI. This work investigates users' mental models of an AI-decision aid. An experiment was designed to mimic a realistic emergency preparedness scenario in which a resource must be allocated into 1 of 100 possible locations based on a variety of dynamic visual heat maps. The users are assisted in resource placement by an AI-decision aid. The experiment was divide into two experimental blocks. The first of which was used to determine mental model accuracy. The second of which was used to examine preferences in meeting the individual and human-AI team goals. The users are asked to determine whether the AI is satisfying a set of constraints to best serve the affected population. Users are also asked to provide an overall score for the AI performance in resource placement. It was found that users tended to exhibit a binary bias in which they tended to categorize performance into discrete bins rather than on a continuous scale, however, the users were able to distinguish between individual and team goals in a human-AI team decision task and did not exhibit a bias towards the human goals. Mental Models, Decision Support, Artificial Intelligence

Metrics for human-robot teaming should extend to teams consisting of multiple human and robotic agents, and to teams working in complex, dynamic work domains. This work proposes that to comprehensively analyze and evaluate multi-human, multi-robot teams, traditional HRI metrics of performance, and efficiency must be expanded upon to incorporate metrics of teamwork. We develop five distinct metrics to capture both ecological and cognitive aspects of teamwork found to be important in human-automation interaction, inspired by research in the cognitive systems engineering (CSE) community. We demonstrate the application of these metrics in a spacecraft maintenance case study comparing multiple human-robot team architectures. The case study demonstrates that the teamwork metrics capture aspects of human-robot interaction (HRI) not apparent when using only traditional performance and efficiency metrics. The article concludes that the proposed teamwork metrics are complementary to existing metrics in HRI and should be included in the evaluation of human-robot teams.

As automation becomes more prevalent, the fear of job loss due to automation increases [22]. Workers may not be amenable to working with a robotic co-worker due to a negative perception of the technology. The attitudes of workers towards automation are influenced by a variety of complex and multi-faceted factors such as intention to use, perceived usefulness and other external variables [15]. In an analog manufacturing environment, we explore how these various factors influence an individual's willingness to work with a robot over a human co-worker in a collaborative Lego building task. We specifically explore how this willingness is affected by: 1) the level of social rapport established between the individual and his or her human co-worker, 2) the anthropomorphic qualities of the robot, and 3) factors including trust, fluency and personality traits. Our results show that a participant's willingness to work with automation decreased due to lower perceived team fluency (p=0.045), rapport established between a participant and their co-worker (p=0.003), the gender of the participant being male (p=0.041), and a higher inherent trust in people (p=0.018).

Decision makers frequently encounter environments without perfect information, in which factors such as the distribution of missing information and estimates of missing information significantly impact decision accuracy and speed. This work presents an experiment which modifies an environment with missing information (total information, option imbalance, cue balance) and examines user estimates of the missing information to understand how accuracy and decision speed respond under time pressure. Results indicate that regardless of the way missing information is estimated, certain distributions of missing information reduce decision accuracy. Results from this work also indicate that beyond information distribution and estimation strategy, differences in decision strategy adopted may explain significant differences in decision performance. High performers tend to ignore a greater percentage of information instead of attempting to estimate it, thereby adopting a strategy more heuristic in nature.

Recently, research by groups in academia, industry, and government has shifted toward the development of AI and machine learning tools to advise human decision-making in complex, dynamic problems. Within this collaborative environment, humans alone are burdened with the task of managing team strategy due to the AI-agent’s use of an unrealistic model of the human-agent’s decision-making process. This work investigates the use of an unsupervised machine learning method to enable AI-systems to differentiate between human decision-making strategies, enabling improved team collaboration and decision support. An interactive experiment is designed in which human-agents are subjected to a complex decision-making environment (a storm tracking interface) in which the provided visual data sources change over time. Behavioral data from the human-agent is collected, and a k-means clustering algorithm is used to identify individual decision strategies. This approach provides evidence of three distinct decision strategies which demonstrated similar degrees of success as measured by task performance. One cluster utilized a more analytic approach to decision-making, spending more time observing and interacting with each data source, while the other two clusters utilized more heuristic decision-making strategies. These findings indicate that if AI-based decision support systems utilize this approach to distinguish between the human-agent’s decision strategies in real time, the AI could develop an improved “awareness” of team strategy, enabling better collaboration with human teammates.

Visual engagement is common in many situations where human operators must perform tasks in challenging environments. This visual engagement has the potential to impact the safety of these operators when dealing with dynamic threats. Perceptual cues have been shown to elicit physical evasion maneuvers, thereby improving safety. In this article, we investigate the effects of cues and visual engagement on rapid whole-body responses. The visual task, inspired by the Trail Making Test (TMT), served as a proxy for visual engagement in the real world. Our continuous TMT minigame and threat simulation are implemented in a virtual reality environment. Participants attempt to maximize their performance score by quickly solving TMTs and dodging dynamic threats from various in-plane directions. They are provided with no cues (control), visual cues, and vibrotactile cues indicating impending threat directions. Participant’s ability to dodge threats is quantified by failure rate and reaction time within field of view and for all approach directions. An index of difficulty highlighted perceptual cue response sensitivity to varying threat speeds and sizes. This article provides two core key contributions and other interesting findings: 1) the results illustrate that tactile cues enable statistically significantly better dodging rates than visual cues or with human vision alone (control condition); and 2) that visual engagement degrades human evasion performance in a statistically significant way. Finally, tactile cue responses appear to be less sensitive than visual cues to visually engaging tasks within the higher portion of difficulty index range that is investigated.

Embodiment of actions and tasks has typically been analyzed from the robot's perspective where the robot's embodiment helps develop and maintain trust. However, we ask a similar question looking at the interaction from the human perspective. Embodied cognition has been shown in the cognitive science literature to produce increased social empathy and cooperation. To understand how human embodiment can help develop and increase trust in human-robot interactions, we created conducted a study where participants were tasked with memorizing greek letters associated with dance motions with the help of a humanoid robot. Participants either performed the dance motion or utilized a touch screen during the interaction. The results showed that participants' trust in the robot increased at a higher rate during human embodiment of motions as opposed to utilizing a touch screen device.

A goal of interactive machine learning (IML) is to enable people with no machine learning knowledge to intuitively teach intelligent agents how to perform tasks. This study investigates how three factors of the design of an interactive reinforcement learning agent - generalization through time, immediacy, and a time delay - impact the user's experience with the agent. We conducted a human-subject experiment in which people trained four agents with different interaction designs to play a simple game using verbal instruction. All agent variations were modified versions of the Newtonian Action Advice algorithm, an interactive reinforcement learning agent that learns from verbal advice like, "go left.'' The results show that both a time delay and probabilistic interface created poor user experiences. This is particularly important for IML designers, because the current algorithms almost universally are probabilistic and do not immediately respond to the human's input.

A goal of interactive machine learning (IML) is to enable people with no specialized training to intuitively teach intelligent agents how to perform tasks. Toward achieving that goal, we are studying how the design of the interaction method for a Bayesian Q-Learning algorithm impacts aspects of the human’s experience of teaching the agent using human-centric metrics such as frustration in addition to traditional ML performance metrics. This study investigated two methods of natural language instruction: critique and action advice. We conducted a human-in-the-loop experiment in which people trained two agents with different teaching methods but, unknown to each participant, the same underlying reinforcement learning algorithm. The results show an agent that learns from action advice creates a better user experience compared to an agent that learns from binary critique in terms of frustration, perceived performance, transparency, immediacy, and perceived intelligence. We identified nine main characteristics of an IML algorithm’s design that impact the human’s experience with the agent, including using human instructions about the future, compliance with input, empowerment, transparency, immediacy, a deterministic interaction, the complexity of the instructions, accuracy of the speech recognition software, and the robust and flexible nature of the interaction algorithm.

Effective human-aware robots should anticipate their user’s intentions. During hand-eye coordination tasks, gaze often precedes hand motion and can serve as a powerful predictor for intent. However, cooperative tasks where a semi-autonomous robot serves as an extension of the human hand have rarely been studied in the context of hand-eye coordination. We hypothesize that accounting for anticipatory eye movements in addition to the movements of the robot will improve intent estimation. This research compares the application of various machine learning methods to intent prediction from gaze tracking data during robotic hand-eye coordination tasks. We found that with proper feature selection, accuracies exceeding 94% and AUC greater than 91% are achievable with several classification algorithms but that anticipatory gaze data did not improve intent prediction.

The research question addressed by this study was: What information should be presented to or hidden from decision makers in order to facilitate high performance in decision tasks? Previous research on information search is limited because of its focus on analytic information acquisition methods; analytic because of the focus on maximizing expected utility; acquisition because of the focus on what information should be added or searched for. Implementing these methods requires reliable assessments of probabilities, cue weights, and cue values and does not provide suggestions on how to restrict or remove information. In this work, we present four heuristics, or simple rules, for acquiring and restricting information that only require an understanding of the distribution of known and unknown information (information imbalance and complete attribute pairs). The rules were tested on a range of analytic and heuristic decision strategies within two-option decision tasks across 15 real-world environments. Though the rules are transparent and easy to communicate (create a balance of information between options and within cues) and require little information to perform, the simulation results show that the rules were generally effective across all environments. For almost every combination of rule and strategy, the heuristic restriction rules were shown to be more likely to increase rather than decrease accuracy. In every combination, the heuristic acquisition rules were shown to increase accuracy more than acquiring information that did not adhere to the rules. Further statistical and mathematical analysis showed that rules are mediated by strategies' full information accuracy and estimates of missing information.

Decision makers are often required to make decisions with incomplete information. In order to design decision support systems (DSSs) utilizing restrictiveness and guidance to assist decision makers in these situations, it is essential to understand how certain decision-making strategies are affected by incomplete information. This paper presents the results of a simulation measuring the accuracy and effort of two heuristic strategies, take-the-best and Tallying, alongside two analytic decision-making strategies, weighted-additive and equal-weighting, in scenarios with varying levels of total information, information imbalance, dispersion, and dominance. Correct decisions were determined by the option with the higher overall score from the weighted-additive model with full information. Effort was measured as counts of elementary information processes required by each strategy to make decisions. Multi- and one-way statistical analyses measured the effect of total information, information imbalance, dispersion, and dominance, on accuracy and effort required for each decision strategy. Three principle results were found: 1) context features matching naturalistic decision settings result in heuristic strategies being closest in accuracy to analytic strategies; 2) the variability in the distribution of the effort requirements of the heuristic strategies for each level of total information indicates that the effort requirements of heuristics may not always be as favorable as prior studies have shown; and 3) the tradeoff between information imbalance and total information suggests new insight for DSS design of restrictiveness and guidance for scenarios with incomplete information.

Ecological rationality is the study of when certain decision making strategies exploit specific environments so efficiently that further information and computation would not necessarily increase accuracy. This perspective challenges many of the normative accounts of rationality by arguing that heuristics, i.e. decision making strategies that ignore some available information, in some environments can make decisions faster, more efficiently, and/or more accurately than analytic decision making strategies. The challenge for many researchers and proponents of ecological rationality is to determine when heuristics are ecologically rational. Research has previously identified some important environmental structures: uncertainty, redundancy, sample size, and variability in weights. This paper introduces a new environmental structure, a measure of distribution of incomplete information called complete attribute pairs, and shows how this structure indicates when two well-studied heuristics, Take-the-Best (TTB) and Tallying, are ecologically rational under conditions of incomplete information. Specifically this paper presents the results of a simulation measuring the accuracy and effort of TTB and Tallying alongside two analytic decision making strategies, weighted-additive (WADD) and equal weighting (EW), in scenarios with incomplete information, showing that the analytic strategies were almost invariant to changes in complete attribute pairs while increases in complete attribute pairs increased the accuracy of TTB and Tallying. These results identify complete attribute pairs as a parameter that is potentially capable of indicating ecological rationality of TTB and Tallying.

In this paper we investigate the use of Work Domain Analysis (WDA), a technique from the field of cognitive engineering, to inform the creation of options and constraints for Reinforcement Learning (RL) algorithms. The micro-world of Pac-Man, a classic arcade game, is used as a tractable and representative work domain. WDA was conducted on individuals familiar with Pac-Man and an Abstraction Hierarchy (AH), a means-ends representation of their understanding of the game, was created for each individual. The abstraction hierarchies for best performing and worst performing individuals were then combined to illustrate the differences between the different groups. Several differences between the two groups were found, and included the use of defense as well as offensive strategies by high performers versus only defense by poor performers, context sensitivity and additional goals and more sophisticated constraints by high performers. The differences were translated into an options and constraint paradigm suitable for incorporation into RL algorithms.

Aircraft automation designers are faced with the challenge to develop and improve automation such that it is transparent to the pilots using it. To identify problems that may arise between pilots and automation, methods are needed that can uncover potential problems with automation early in the design process. In this paper, simulation and model checking are combined and their respective advantages leveraged to find problematic human-automation interaction using methods that would be available early in the design process. A particular problem of interest is automation surprises, which describe events when pilots are surprised by the actions of the automation. The Tarom flight 381 incident involving the former Airbus automatic speed protection logic, leading to an automation surprise, is used as a common case study. Results of this case study indicate that both methods identified the automation surprise found in the Tarom flight 381 incident, and that the simulation identified additional automation surprises associated with that flight logic. The work shows that the methods can be symbiotically combined, and the joint method is suitable to identify problematic human-automation interaction such as automation surprise.

To identify problems that may arise between pilots and automation, methods are needed that can uncover potential problems with automation early in the design process. Such potential problems include automation surprises, which describe events when pilots are surprised by the actions of the automation. In this work, agent-based, hybrid time simulation and model checking are combined and their respective advantages leveraged in an original manner to find problematic human-automation interaction (HAI) early in the design process. The Tarom 381 incident involving the former Airbus automatic speed protection logic, leading to an automation surprise, was used as a common case study for both methodology validation and further analysis. Results of this case study show why model checking alone has difficulty analyzing such systems and how the incorporation of simulation can be used in a complementary fashion. The results indicate that the method is suitable to examine problematic HAI, such as automation surprises, allowing automation designers to improve their design.

This paper explores methods to include aspects of controller taskload into conflict-resolution programs through a parametric approach. We are motivated by the desire to create conflict-resolution decision-support tools that operate within a human-in-the-loop control architecture by actively accounting for, and moderating, controller taskload. Specifically, we introduce two conflict-resolution programs with the objective of managing controller conflict-resolution taskload, i.e., the number of maneuvers used to separate air traffic. Managing conflict-resolution taskload is accomplished by penalizing aircraft maneuvers through theirL1norm in the cost function or constraining the number of maneuvers directly. Analysis of the programs reveals that both approaches are successful at managing controller conflict-resolution taskload and minimizing fuel burn. Directly constraining conflict-resolution taskload is more successful at minimizing the variation in the number of aircraft maneuvers issued and returning the aircraft to their desired exit point. Penalizing maneuvers throughL1norm costs is more successful at reducing controller conflict-resolution taskload at lower traffic volumes. Ultimately, results demonstrate that the inclusion of such parametric models can successfully regulate controller conflict-resolution taskload.

Breakdowns in complex systems often occur as a result of system elements interacting in ways unanticipated by analysts or designers. The use of task behavior as part of a larger, formal system model is potentially useful for analyzing such problems because it allows the ramifications of different human behaviors to be verified in relation to other aspects of the system. A component of task behavior largely overlooked to date is the role of human-human interaction, particularly human-human communication in complex human-computer systems. We are developing a multi-method approach based on extending the Enhanced Operator Function Model language to address human agent communications (EOFMC). This approach includes analyses via theorem proving and future support for model checking linked through the EOFMC top level XML description. Herein, we consider an aviation scenario in which an air traffic controller needs a flight crew to change the heading for spacing. Although this example, at first glance, seems to be one simple task, on closer inspection we find that it involves local human-human communication, remote human-human communication, multi-party communications, communication protocols, and human-automation interaction. We show how all these varied communications can be handled within the context of EOFMC.