There is significant interest in enabling robots to learn to perform tasks directly from interactions with non-expert users. Typically, a human serves as a teacher whose only task is to provide feedback to a robot learner. However, in real-world human-robot collaborations, the human often assists with the task while also offering feedback. Our key insight is that we can extract additional, implicit feedback from the human’s actions during the collaboration to augment the robot learning process. Under the assumption of fixed-role assignments, we first propose to formalize human preferences over a human-robot collaboration as a shared set of parameters encoding alignment between two reward functions: one that drives human behavior, and another that should direct robot behavior. This allows us to extract implicit feedback from an interaction by reasoning about the human’s actions in the task as actions that reveal the human’s preferences. Then, we combine this implicit feedback with traditional explicit human feedback to facilitate estimating the human’s preferences. We evaluated our proposed approach for Preference learning from Implicit and Explicit feedback (PIE) in simulations and with real users in a cooking scenario. Our simulation results indicate that combining multiple modalities of human feedback improves a robot’s ability to estimate human preferences over the collaboration, with a similar trend observed in real-world evaluations. These findings highlight a promising direction for enabling robots to adapt to a user’s preference model more quickly, thereby reducing the amount of time a person must spend teaching a robot.

Fairness is critical for collaboration between humans, and recent research has shown its importance in human–robot collaboration. However, most human-robot interaction (HRI) studies probe fairness judgments toward a robot only at the conclusion of an interaction, overlooking the fact that perceptions of fairness can evolve over time. We present two studies of dynamic fairness that both leverage a Multiplayer Space Invaders game, where a robot controls a spaceship and distributes support across players' sides of the screen. The robot's support is at times biased in favor of one player or the other. In the first study, we examine how fairness perceptions are influenced by the timing of a robot's biased support (early vs. late in the interaction) and the beneficiary of this support (the participant vs. another agent). In the second study, we investigate how expectations of a robot's support behavior (biased vs. unbiased) interact with its actual behavior (biased vs. unbiased) in a setting where two participants each worked to score an individual score threshold. We find that fairness judgments are dynamic: fairness falls after the robot's allocation of support becomes biased but is slower to recover once support becomes unbiased, and participants expecting unbiased behavior judge fairness more harshly when these expectations are violated. Our findings advance understanding of fairness in HRI by presenting it as a dynamic construct shaped not only by the actual behavior of the robot but also by the timing of robot actions and expectations of robot behavior.

My work aims to enable robots to better learn from human feedback in human-robot interactions. The way in which people want to collaborate with a robot can vary person-to-person, interaction-to-interaction, or even within an interaction with a given person. Thus, robots need to be able to adapt their behavior during interactions. Robots typically learn from humans via explicit feedback, such as evaluative feedback, preferences, or demonstrations. We know that humans also provide additional information implicitly through non-verbal behavior that gives clues about their internal states during interactions. My work investigates how we can incorporate both kinds of feedback into robot learning paradigms.

The rapid and nearly pervasive impact of artificial intelligence on fields as diverse as medicine, law, banking, and the arts has made many students who would never enroll in a computer science class become interested in understanding elements of artificial intelligence. Fueled by questions about how this technology would change their own fields, these students are not seeking to become experts in building AI systems but instead are searching for a sufficient understanding to be safe, effective, and informed users. In this paper, we describe a first-of-its-kind course offering, "Artificial Intelligence for Future Presidents" designed and taught during the spring of 2024. We share rationale on the design and structure of the course, consider how best to convey complex technical information to students without the background in programming or mathematics, and consider methods for supporting an understanding of the limits of this technology.

Recent work in Human-Robot Interaction (HRI) has shown that robots can leverage implicit communicative signals from users to understand how they are being perceived during interactions. For example, these signals can be gaze patterns, facial expressions, or body motions that reflect internal human states. To facilitate future research in this direction, we contribute the \textttREACT database, a collection of two datasets of human-robot interactions that display users' natural reactions to robots during a collaborative game and a photography scenario. Further, we analyze the datasets to show that interaction history is an important factor that can influence human reactions to robots. As a result, we believe that future models for interpreting implicit feedback in HRI should explicitly account for this history. \textttREACT opens up doors to this possibility in the future.

One important aspect of effective human--robot collaborations is the ability for robots to adapt quickly to the needs of humans. While techniques like deep reinforcement learning have demonstrated success as sophisticated tools for learning robot policies, the fluency of human-robot collaborations is often limited by these policies' inability to integrate changes to a user's preferences for the task. To address these shortcomings, we propose a novel approach that can modify learned policies at execution time via symbolic if-this-then-that rules corresponding to a modular and superimposable set of low-level constraints on the robot's policy. These rules, which we call Transparent Matrix Overlays, function not only as succinct and explainable descriptions of the robot's current strategy but also as an interface by which a human collaborator can easily alter a robot's policy via verbal commands. We demonstrate the efficacy of this approach on a series of proof-of-concept cooking tasks performed in simulation and on a physical robot.

Humans expect robots to learn from their feedback and adapt to their preferences. However, there are limitations with how humans provide feedback to robots, e.g., humans may give less feedback as interactions progress. Therefore, it would be advantageous if robots could influence humans to provide more feedback during interactions. We conducted a 2x2 between-subjects user study (N=71) to investigate whether the framing and timing of a robot's reminder to provide feedback could influence human interactants. Human-robot interactions took place in the context of Space Invaders, a fast-paced and continuous collaborative environment. Our results suggest that reminders can influence the amount of feedback humans provide to robots, how participants feel about the robot, and how they feel about providing feedback during the interaction.

Recent research in robot learning suggests that implicit human feedback is a low-cost approach to improving robot behavior without the typical teaching burden on users. Because implicit feedback can be difficult to interpret, though, we study different methods to collect fine-grained labels from users about robot performance across multiple dimensions, which can then serve to map implicit human feedback to performance values. In particular, we focused on understanding the effects of annotation order and frequency on human perceptions of the self-annotation process and the usefulness of the labels for creating data-driven models to reason about implicit feedback. Our results demonstrate that different annotation methods can influence perceived memory burden, annotation difficulty, and overall annotation time. Based on our findings, we conclude with recommendations to create future implicit feedback datasets in Human-Robot Interaction.

Much prior work on creating social agents that assist users relies on preconceived assumptions of what it means to be helpful. For example, it is common to assume that a helpful agent just assists with achieving a user’s objective. However, as assistive agents become more widespread, human-agent interactions may be more ad-hoc, providing opportunities for unexpected agent assistance. How would this affect human notions of an agent’s helpfulness? To investigate this question, we conducted an exploratory study (N=186) where participants interacted with agents displaying unexpected, assistive behaviors in a Space Invaders game and we studied factors that may influence perceived helpfulness in these interactions. Our results challenge the idea that human perceptions of the helpfulness of unexpected agent assistance can be derived from a universal, objective definition of help. Also, humans will reciprocate unexpected assistance, but might not always consider that they are in fact helping an agent. Based on our findings, we recommend considering personalization and adaptation when designing future assistive behaviors for prosocial agents that may try to help users in unexpected situations.

While neural network binary classifiers are often evaluated on metrics such as Accuracy and $F_1$-Score, they are commonly trained with a cross-entropy objective. How can this training-evaluation gap be addressed? While specific techniques have been adopted to optimize certain confusion matrix based metrics, it is challenging or impossible in some cases to generalize the techniques to other metrics. Adversarial learning approaches have also been proposed to optimize networks via confusion matrix based metrics, but they tend to be much slower than common training methods. In this work, we propose a unifying approach to training neural network binary classifiers that combines a differentiable approximation of the Heaviside function with a probabilistic view of the typical confusion matrix values using soft sets. Our theoretical analysis shows the benefit of using our method to optimize for a given evaluation metric, such as $F_1$-Score, with soft sets, and our extensive experiments show the effectiveness of our approach in several domains.