Source-linked AI summary
What Questions Should Robots Be Able to Answer? A Dataset of User Questions for Explainable Robotics
Lennart Wachowiak, Andrew Coles, Gerard Canal, Oya Celiktutan
TL;DR
Explainable robotics lacks a broad empirical account of what users want to ask household robots, a gap relevant to understanding their information needs. This paper introduces a dataset of user questions and finds that questions about hypothetical difficulties and correct behavior rank highest in importance.
Problem
Despite growing interest in explainable robotics, there is no broad empirical account of what users want to ask robots across everyday tasks.
Source: §1 · p. 2Method
The paper presents a dataset of questions users have for household robots.
Source: §1 · p. 2Results
Questions about how robots handle hypothetical scenarios and ensure correct behavior rank highest in importance.
Source: §1 · p. 2Takeaways & Limitations
Users may want to understand how a robot works, makes decisions, and handles unexpected issues.
Source: §1 · p. 1The elicited questions reflect reactions to varied scenarios, and the authors seek an even broader, more diverse set of questions.
Source: §5.4 · p. 20
Abstract
from arXiv · showhide
With the growing use of large language models and conversational interfaces in human-robot interaction, robots' ability to answer user questions is more important than ever. We therefore introduce a dataset of 1,893 user questions for household robots, collected from 100 participants and organized into 12 categories and 70 subcategories. Most work in explainable robotics focuses on why-questions. In contrast, our dataset provides a wide variety of questions, from questions about simple execution details to questions about how the robot would act in hypothetical scenarios -- thus giving roboticists valuable insights into what questions their robot needs to be able to answer. To collect the dataset, we created 15 video stimuli and 7 text stimuli, depicting robots performing varied household tasks. We then asked participants on Prolific what questions they would want to ask the robot in each portrayed situation. In the final dataset, the most frequent categories are questions about task execution details (21.4%), the robot's capabilities (12.6%), and performance assessments (10.7%). Although questions about how robots would handle potentially difficult scenarios and ensure correct behavior are less frequent, users rank them as the most important for robots to be able to answer. Moreover, we find that users who identify as novices in robotics ask different questions than more experienced users. Novices are more likely to inquire about simple facts, such as what the robot did or the current state of the environment. As robots enter environments shared with humans and language becomes central to giving instructions and interaction, this dataset provides a valuable foundation for (i) identifying the information robots need to log and expose to conversational interfaces, (ii) benchmarking question-answering modules, and (iii) designing explanation strategies that align with user expectations.
1 Introduction
The paper presents a broad empirical account of questions people want household robots to answer, addressing a gap in explainable robotics. Its findings show that users especially value answers about difficult hypothetical situations and correct behavior, with question types also varying by robot experience.
1 Introduction
- The dataset addresses the lack of a broad empirical account of questions users have for robots across everyday tasks.The authors focus on general-purpose robots performing household tasks in settings involving close human contact and natural-language communication.Source: §1 · p. 2
- The dataset can help roboticists anticipate user questions, identify information to log and expose to language interfaces, and support question-answering and XAI benchmarks.Examples of relevant information include action histories, object states, capability boundaries, and sources of knowledge.Source: §1 · p. 2
- 1,893 user questions are organized into 12 categories and 70 subcategories, spanning topics beyond the causal why-questions common in explainable robotics.The categories include execution details, robot capabilities, potential difficulties, technical details, and sources of knowledge.Source: §1 · p. 2
- Questions about hypothetical scenarios and how robots ensure correct behavior rank highest in importance, while why-questions have the second-lowest average importance.Questions about safety, failure, and potential issues form a unifying theme among highly ranked questions.Source: §1 · p. 2
- Novices ask relatively more about execution details and environmental state, while experienced users more often ask about difficult situations and robot decisions.Users with more positive attitudes toward robots also consider question-answering ability more important.Source: §1 · pp. 2–3
2 Related Work
The related work situates the study among explainability research, user-difference studies, HRI datasets, and question taxonomies. It distinguishes this work through its focus on concrete user questions about robots performing tasks, rather than other interaction types or social small talk.
2 Related Work
- The section reviews explainable robotics, variation in users’ explanation needs, HRI datasets, and prior categorizations of questions for AI and robots.It frames the paper against work on natural-language question answering and LLMs as well as earlier dataset and taxonomy efforts.Source: §2 · p. 3
2.1 Explainable AI/Robotics
- Robot question-answering systems rely on information available to the LLM, making users’ questions relevant to what systems should record and expose.The paper positions its dataset as a guide to users’ information needs and to data available to downstream question-answering modules.Source: §2.1 · p. 3
2.2 User Differences in Explanation Needs
- Prior work motivates examining user differences, including robot experience and attitudes, because explanation preferences may vary across people.The authors explore whether people with more positive attitudes toward robots consider question-answering more important.Source: §2.2 · pp. 3–4
2.3 Datasets for HRI
- Many HRI datasets cover instructions, social cues, or robot clarification questions, but are not designed to analyze questions users ask robots.Some dialogue corpora instead record people instructing robots or human–human collaboration on household tasks.Source: §2.3 · p. 4
2.4 Question Categorizations
- This study elicits task-context questions about a robot executing a specific task, rather than social small talk, and offers a larger, XAI-relevant question dataset.The authors contrast it with prior work on open-day interactions and other question categorizations that do not provide this kind of task-linked natural-language dataset.Source: §2.4 · pp. 4–5
3 Method — How the Dataset was Collected
This section describes how questions about robots’ household-task behavior were elicited across video and text scenarios, organized into a two-level coding scheme, and analyzed. It also notes interpretive limits for ambiguous questions and outlines the use of example answers, mixed-effects models, and user-characteristic analyses.
3.1 Study Procedure
- Participants saw four or five randomized videos and two text scenarios, supplied two scenario-specific and two broader questions, and rated question importance on a 1–5 scale.Source: §3.1 · p. 5
3.2 Video and Text Stimuli
- Stimuli spanned household tasks, humanoids and mobile manipulators, and solo, human, or multi-robot settings; seven text reports included occasional failures.Videos averaged 35.4 seconds and were cut to show one self-contained task.Source: §3.2 · pp. 5–6
3.3 Participants
- The final dataset contains 1,893 questions from 100 participants, excluding attention-check failures and nonsensical free-text responses.Participants were fluent in English and had a 100% Prolific approval rate with at least five prior studies.Source: §3.3 · p. 6
3.4 Methods for Data Analysis
- Questions were coded inductively and bottom-up, with descriptive codes iteratively merged into a two-level hierarchy and reviewed in a second pass.The authors used manual annotation because existing categorizations did not capture the range of questions and automated clustering can miss fine-grained user intent.Source: §3.4 · p. 8
- On 96 balanced re-annotated questions, annotators agreed on 79% (Cohen’s kappa = .77, CI = [.68, .86], p < .0001), classified as “substantial agreement” by Landis et al..The most common disagreement, occurring three times, was between self/task-assessment and execution-details, so the codebook definitions were clarified.Source: §3.4 · p. 9
- Ambiguous questions can have multiple interpretations, and disagreements can only be resolved by asking the participant who provided the question.Source: §3.4 · p. 9
- Future dataset extensions could ask participants to provide example answers they would find satisfying, helping annotators understand their expectations.Source: §3.4 · p. 9
- For RQ2, a linear mixed-effects model accounts for repeated, non-independent ratings by modeling participant and stimulus identifiers as random effects.Source: §3.4 · p. 9
- The RQ2 model uses question category as the fixed effect and importance score as the outcome, is fit by restricted maximum likelihood, and applies Holm–Bonferroni correction.Source: §3.4 · p. 9
- For RQ3, the authors use Spearman rank correlations to examine relationships between user characteristics and question-importance scores across categories.Source: §3.4 · p. 9
- They also use multiple linear regression, with user attitudes and robotics experience as predictors and question importance as the outcome.Source: §3.4 · p. 9
4 Results — A Dataset of Questions for a Robot
The results characterize diverse household-robot questions by category and language, including how they were curated, with execution details the most common category and a long tail of rarer types. They also report that users rate questions about potential issues as most important and describe differences in question patterns associated with experience and robot attitudes, including more questions in four categories from participants with high robot experience.
4 Results — A Dataset of Questions for a Robot
- Of 2,037 collected questions, 144 were excluded, mostly commands phrased as questions or unclear, incomplete, or joking entries; participant answers and analysis code are publicly available.Participants contributed an average of 18.93 questions (SD=2.82), and the average question was 9.73 words long (SD=4.16).Source: §4 · p. 10
- Questions vary in perspective: 67% address the robot directly, 30% refer to it or the environment in third person, and 2% use a first-person perspective.Some questions depend on the specific scenario for their meaning.Source: §4 · p. 10
4.1 Question Categories
- Execution-detail questions are the most common, often asking what happened during task execution, while a long tail of rarer question types complicates reliable question answering across situations.The dataset includes questions about capabilities, difficulty, correctness, hypothetical challenges, decision-making, and other themes.Source: §4 · p. 12, §4.1 · p. 13
- Question categories cover capabilities, task success and difficulty, environment state, potential problems and safeguards, and the robot’s knowledge, mental states, and decisions.Other categories include execution details, timing, future actions, and technical details such as algorithms, perception, settings, and hardware.Source: §4.1 · pp. 13, 15
- Why-questions often request action reasons or contrasts, with explicit contrasts in 28% of them; users may also expect inferred alternatives or explanations that help predict future actions.The category also includes questions about knowledge sources such as perception, learning, and prior information, and about the procedures underlying decisions.Source: §4.1 · pp. 13, 15
4.2 Question Importance Scores
- Potential-issues questions ranked highest (EMM = 4.01, 95% CI = [3.82, 4.20], n = 118), versus why (3.58, [3.42, 3.75], n = 190) and mental-state (3.53, [3.35, 3.72], n = 120) questions.Potential-issues questions concern difficult task variations and how the robot can act correctly and avoid problems; potential-issues ratings exceeded why ratings by 0.43 EMM points (p = .0007).Source: §4.2 · p. 16, Figure 3 · p. 16
4.3 Differences between Participant Groups
- Importance ratings correlated with personal-level (ρ = .33, p = .0037) and societal-level robot attitudes (ρ = .32, p = 0.0047), but not significantly with robotics experience (ρ = .19, p = .154).A regression model was significant overall (F(3, 96) = 5.26, p = .002) and explained 14.1% of importance-score variance; personal-level attitudes were the strongest predictor (coefficient 0.14, CI = [0.014, 0.272]).Source: §4.3 · pp. 17–18
- Participants with low robotics experience asked execution-details questions 23.3% of the time, compared with 15.7% among more experienced participants (difference = -7.6%).The paper also reports a similar contrast for environment-state questions, but the supplied passage gives no values for that comparison.Source: §4.3 · p. 18, Figure 5 · p. 18
- The cited passage lists four categories with positive changes: potential-issues (+4.6%), how-did-you-decide (+3.5%), mental-state (+3.1%), and what-abilities (+2.9%).Source: §4.3 · p. 18
5 Discussion and Conclusion
The discussion reviews the dataset’s question categories, importance ratings, and differences between participant groups, then outlines limitations that bound its interpretation. It highlights how these findings can inform robot question-answering design.
5.1 Question Categories
- The categories can guide what information robots should collect, including environmental state, capabilities, and task performance.The discussion notes that answering environmental-state questions may require recording data during task execution.Source: §5.1 · pp. 18–19
- Text scenarios elicited more execution-detail questions than videos (36% vs. 16%), while videos elicited more questions about abilities and mental state.The authors relate the difference to text providing only a high-level overview while videos show execution details.Source: §5.1 · p. 19
5.2 Question Importance Scores
- Potential-issue questions rank highest in importance, while why-questions and questions about robot mental states rank low.No category scored below 3.5 on the 1–5 scale.Source: §5.2 · p. 19
5.3 Differences between Participant Groups
- Participants with more positive attitudes toward robots assigned greater average importance to their questions, while novices asked more execution-detail and environment-state questions.Novices’ questions in those categories concerned simple facts about what happened and the current state of things.Source: §5.3 · p. 19
5.4 Limitations and Future Work
- Some questions were difficult to interpret, and the resulting categories were necessarily shaped by the main annotator.The authors suggest collecting users’ example answers as one possible way to address uncertainty about intended meaning.Source: §5.4 · p. 20
- Text and video stimuli have lower ecological validity than real-world interactions, although they enable controlled, reproducible testing across diverse scenarios.Participants saw each specific task by a specific robot only once, so the study does not trace questions across repeated interactions.Source: §5.4 · p. 20
- The sample mostly came from Western and Anglophone countries, and the analysis treats household-robot questions as task-agnostic.The authors identify cultural and task-specific analyses as directions for follow-up work.Source: §5.4 · pp. 20–21
A Online Resources
The appendix identifies where the research materials and data are available. It lists the dataset, annotations, and questionnaire among the repository contents.
A Online Resources
- The GitHub repository contains the natural-language questions, their categories and importance scores, second-annotator annotations, and the questionnaire.The questionnaire is provided to support reproducibility.Source: A Online Resources · pp. 23–24
- The repository includes second-annotator annotations for computing inter-annotator agreement.Source: A Online Resources · p. 24
- The dataset is also hosted on Hugging Face.Source: A Online Resources · p. 24
B Video Stimuli References
The appendix documents the video sources used for the stimuli and notes how selected clips were prepared. It also identifies the fair-use basis stated by the authors.
B Video Stimuli References
- The video-stimulus references identify source videos and the household tasks selected from them.When a video contains multiple tasks, the appendix specifies the timestamp of the task used.Source: B Video Stimuli References · p. 24
- Stimulus preparation sometimes involved changing playback speed or cropping out teleoperators and text boxes.The authors say to contact the corresponding author to see the exact stimuli used.Source: B Video Stimuli References · p. 24
- The authors state that their use of the videos falls under YouTube’s fair-use policy for research, and that the table frames show no identifiable people.Source: B Video Stimuli References · p. 24