Highlights
Thank you to all the participants for their contributions to the Imperfectly Relatable Robot 2023 workshop, which was held as a companion to Human Robot Interaction Conference. We had an overwhelming response both in terms of participation and paper contributions demonstrating that failures are a growing and interesting research topic within the HRI community. We were very pleased and honoured to have an excellent keynote and panel speakers.
Below we have summarized the highlights of different events that happened during the workshop, together with suggestions for future directions that could be explored at follow-up events.
Keynote: Disobedience and Intended Failures in HRI by Reuth Mirsky
What can be considered as a failure in HRI? Does it refer to the technical misfunctions of the robot alone? This talk emphasized the different types of failures, such as interaction failures, environmental failures and social norm violations, apart from the technical failures that could occur during a human-robot interaction. “Not all failures are bad” was the theme of the talk.
Intelligent disobedience is a guide dog training method, where if the dog understands that the commands initiated by the handler, could put the handler in danger, it will disobey the handler commands. This behavior is one way of using failures as a useful tool in avoiding mistakes and for safety.
How might interruptions be used for effective human-robot teaming? The idea discussed here involved using robots to ask questions to humans and get help when the given initial instructions appear to be vague or dangerous.
What can be termed as a “good” failure? An intended failure must be reasoned in an understandable way. Designing intended failures to “gracefully” fail is one of the future directions identified.
This talk provided inspiration in how failures might be rethought as opportunities.
Group Discussions:
Workshop participants were invited to propose themes for further discussion in an interactive session in the second part of the session. During this, two themes were identified from a word cloud as being of greatest interest to the participants: explanations and failure criteria.
1. Definitions of failure / failure criteria: The definition of a failure within the HRI context is diverse and varies based on the perception of errors. Cultural aspects also play a part in defining what is considered as a failure event / behavior. Another line of thought that was discussed was, is it ethical to manipulate humans? And when to use failure as a mitigation strategy? Using failures for mitigation is highly context-specific; what could work for one application setting might not work for another.
2. Explanations of Failures: What constitutes a good explanation?
a. conveying certain understanding about certain decisions,
b. what the human thinks the reason could be for the failure,
c. why this happened – reverse explanations can also be used to make the robot learn how and when explanations are needed.
It was noted that over explanation might lead to over estimation of trust in robots, and also that explanations are difficult for a non-verbal robot when compared to the social robots. The success of using failure as a mitigation strategy also depends on people’s expectations towards the robots. We do not expect much from a simple robot but with complex robots we get frustrated if we do not get explanations.
Identified future research directions:
There is a growing interest in using failures as a strategy to prevent and mitigate errors within human-robot context. As the rich discussions above suggest, there are many paths that participants are interested in following. As a growing sub-field, there are extensive improvement opportunities in Identification of failures (How and which event can be termed as a failure event autonomously), Explanations for failures (what is the effective way in explaining failures, how much is good and enough explanations to make the interaction natural), Ethical frameworks for testing failures and different measures and modalities for quantitative and qualitative analysis of failures.