Socially-Aware Robot Explanations: Inferring Needs from Human Facial Expressions

Introduction

Although AI systems can make decisions, their ability to explain them remains limited, particularly in error-prone situations.  This work focuses on mechanistic interpretability in error detection and examines how different explanation types influence user behaviour. While prior research has explored the causes of robot errors,  few have investigated how robots should explain error detection, especially when users express social cues indicating something has gone wrong. We adopt a user-centred approach to transparency, proposing that human facial expressions guide the robot’s explanation behaviour. In this work, we investigate whether users’ facial expressions can inform when and how robots should provide explanations during collaborative tasks. Using a deep-learning model trained on Facial Action Units  from a public HRI dataset,  a robot classified states of user confusion.  

The model was deployed on a robot arm performing a pick-and-place task, where errors were introduced through random perturbations. Based on user behaviour, the robot detected potential errors and provided explanations to increase model transparency. We tested three XAI methods designed to answer how, why, and what-if questions, offering a different aspect of the robot’s decision-making, explaining a robot’s failure detection model. In a study, participants engaged in a robot-assisted pick-and-place task while receiving different types of explanations. User responses were analysed through multimodal signals, alongside subjective measures of cognitive load, trust, and model understanding. The results showed that why-explanations were the most preferred and whatif explanations required more vocal effort. This work demonstrates that facial expressions can be used to tailor explanation frequency demands, supporting transparent and adaptive human-robot interactions.

Method & Results

We used post-hoc, interactive, and model-agnostic explanations,  presented in a consistent format across all methods. Three explanation types were tested using established XAI techniques: How-explanations used global feature importance via SHAP, applying Shapley Value Sampling  to show how features generally influenced the model’s output. Why-explanations provided local feature contributions using Kernel-SHAP, which approximates the decision boundary around a given input (based on the LIME framework ) to highlight input-specific influence.  What-if explanations used counterfactuals to show how minimal changes in two features could alter the model’s prediction, applying the method by Mothilal et al.. 

To stimulate the error detection model, we injected random perturbations into the robot’s input and output, inducing uncertainty. The model, trained on a public dataset of HRI sessions,  used weighted classification and Softmax probabilities within a sliding window to detect user confusion signals.  It traced back to the estimated onset of the error, which served as the input for the explanation algorithms. Participants performed a voice-controlled pickand-place task with the robot arm, and the explanations were displayed on an external monitor. The setup included facial expression and body pose tracking, as well as speech processing.  Explanations were visualised using bar charts, commonly used for tabular data. The robot’s workspace contained PVC pipes, which participants used to assemble various structures.

Several facial expressions were associated with participants’ reactions to different explanation types, with why-explanations eliciting greater expressiveness. What-if explanations, however, were linked to increased vocal effort. For each explanation, we assessed participants’ ability to correctly identify relevant features. We found that how-explanations, based on global feature importance, led to better understanding, likely due to their independence from user-specific input. Overall, participants expressed a clear preference for why-explanations, while what-if explanations were least preferred. Thematic analyses showed that how-explanations were generally seen as clear and easy to follow but lacked personalisation. In contrast, why-explanations were valued for their personalised and user-specific nature, though some found them harder to interpret. Reactions to what-if explanations were mixed: while some participants grasped the counterfactual logic, others found it confusing, reflected in their lower ratings.

Summary & Conclusion

Overall, How-explanations were seen as intuitive and easy to understand, but were perceived as impersonal. Why-explanations were the most preferred and perceived as personalised, though how-explanations led to better objective understanding.

What-if explanations, while expected to be straightforward, required greater vocal effort and elicited lower engagement and expressiveness, despite no increase in reported cognitive load. In summary, our findings in this work offer empirical insights into which explanation types are most effective for error detection models based on facial expressions. We hope this work informs future research on robot error detection by emphasising the importance of multimodal cues and encouraging the integration of error causality with detection, placing multimodal HRI at the centre of XAI-driven interaction design.