Abstract:Machine learning models are commonly tested in-distribution (same dataset); performance almost always drops in out-of-distribution settings. For HRI research, the goal is often to develop generalized models. This makes domain generalization - retaining performance in different settings - a critical issue. In this study, we present a concise analysis of domain generalization in failure detection models trained on human facial expressions. Using two distinct datasets of humans reacting to videos where error occurs, one from a controlled lab setting and another collected online, we trained deep learning models on each dataset. When testing these models on the alternate dataset, we observed a significant performance drop. We reflect on the causes for the observed model behavior and leave recommendations. This work emphasizes the need for HRI research focusing on improving model robustness and real-life applicability.
Abstract:For a robot to repair its own error, it must first know it has made a mistake. One way that people detect errors is from the implicit reactions from bystanders -- their confusion, smirks, or giggles clue us in that something unexpected occurred. To enable robots to detect and act on bystander responses to task failures, we developed a novel method to elicit bystander responses to human and robot errors. Using 46 different stimulus videos featuring a variety of human and machine task failures, we collected a total of 2452 webcam videos of human reactions from 54 participants. To test the viability of the collected data, we used the bystander reaction dataset as input to a deep-learning model, BADNet, to predict failure occurrence. We tested different data labeling methods and learned how they affect model performance, achieving precisions above 90%. We discuss strategies to model bystander reactions and predict failure and how this approach can be used in real-world robotic deployments to detect errors and improve robot performance. As part of this work, we also contribute with the "Bystander Affect Detection" (BAD) dataset of bystander reactions, supporting the development of better prediction models.