CHI Conference on Human Factors in Computing Systems (CHI '26)
Oops, I Did It Again (But I Know It): Robot Failure Consistency and Awareness in Human-Robot Collaboration
Exploring the Impact of Robot Failure Patterns on Trust and Perceived Intelligence
Abstract
In human–robot collaboration, repeated failures are inevitable and can critically impact user trust and perceptions of robot intelligence. This study investigates how different failure types — Freezing, Planning, and Grasping — and the robot’s awareness of these failures influence users across repeated interactions. Using a collaborative Tangram puzzle task, we conducted a mixed-design experiment with 54 participants, manipulating both the failure sequence type (homogeneous vs. heterogeneous) and the robot’s level of failure awareness (none, partial, full). The results reveal that user trust and perceived intelligence are affected by both the current and previous failure, with homogeneous sequences generally causing less trust decline than heterogeneous ones. Robots that demonstrated awareness — either full or partial — were rated significantly higher in trust and perceived intelligence than those showing no awareness, particularly for grasping and planning failures. Qualitative findings further underscore the importance of failure severity and the robot’s recovery behaviour in shaping user perceptions. These insights offer design implications for creating more resilient and trustworthy human–robot interactions amid inevitable errors.
Project Demo
Overview
Robot failures, particularly in real-world environments, pose significant challenges to human-robot collaboration. This research explores the effects of repeated robot failures and the robot's level of awareness on user trust and perceived intelligence. Using a Tangram puzzle task, various failure types were tested with different levels of robot awareness.
Key Contributions
- Identified the influence of failure sequence patterns (homogeneous vs. heterogeneous) on trust and perceived intelligence.
- Demonstrated the benefits of robot awareness in mitigating trust decline following failures.
- Provided insights into designing effective robot behaviors for trust repair in collaborative settings.
Method and System
The study employed a collaborative Tangram puzzle task involving 54 participants. The methodology included manipulating failure sequence types (homogeneous vs. heterogeneous) and robot awareness levels (none, partial, full) to assess their impact on user trust and perceived intelligence.
Results Overview
The experiment revealed that homogeneous failure sequences result in smaller reductions in trust and perceived intelligence compared to heterogeneous sequences. Robots demonstrating awareness were consistently rated higher in trust and perceived intelligence, particularly for grasping and planning failures.
Quantitative Results
Add quantitative results here, such as condition-level metrics, trust score changes, statistical significance tests, or effect sizes.
Qualitative Insights
Add qualitative observations here, such as participant feedback, recovery behaviors, or themes that explain how awareness shaped perception.
Discussion
Limitations
The study was conducted in a controlled laboratory setting, which may limit generalizability to other real-world contexts. The robot's awareness behaviors were pre-scripted, lacking flexibility of real-time autonomous responses.
Future Work
Future research could explore more diverse and severe failure scenarios in real environments. Investigating longer-term user interactions and adaptive repair strategies could significantly advance the field.
Figure Highlights
Citation
@inproceedings{tabatabaei2026robotfailure,
title={Oops, I Did It Again (But I Know It): Robot Failure Consistency and Awareness in Human-Robot Collaboration},
author={Tabatabaei, Ramtin and Kostakos, Vassilis and Johal, Wafa},
booktitle={CHI Conference on Human Factors in Computing Systems},
year={2026}
}