A Multi-Modal Explainability Approach for Human-Aware Robots in Multi-Party Conversation

TL;DR

The paper addresses AE in multi-party HRI and proposes a two-stage approach: an Improved Addressee Estimation (IAE) model for higher accuracy, followed by an Explainable Addressee Estimation (XAE) model that provides intrinsic explanations via attention mechanisms. The XAE is deployed within a modular cognitive architecture on the iCub robot, enabling real-time AE, visualization, and multi-modal explanations (verbal, embodied, graphical). Empirical results show both IAE and XAE outperform the SOTA baseline, with XAE offering interpretable attention maps, modality contribution scores, and time-frame explanations, while user studies indicate strong usefulness and acceptable tolerability of explanations. The work demonstrates that a modular, transparent architecture can enhance human trust and interaction quality in dynamic, multi-party social settings, and highlights avenues for future improvements such as improved diarization and continual learning.

Abstract

The addressee estimation (understanding to whom somebody is talking) is a fundamental task for human activity recognition in multi-party conversation scenarios. Specifically, in the field of human-robot interaction, it becomes even more crucial to enable social robots to participate in such interactive contexts. However, it is usually implemented as a binary classification task, restricting the robot's capability to estimate whether it was addressed \review{or not, which} limits its interactive skills. For a social robot to gain the trust of humans, it is also important to manifest a certain level of transparency and explainability. Explainable artificial intelligence thus plays a significant role in the current machine learning applications and models, to provide explanations for their decisions besides excellent performance. In our work, we a) present an addressee estimation model with improved performance in comparison with the previous state-of-the-art; b) further modify this model to include inherently explainable attention-based segments; c) implement the explainable addressee estimation as part of a modular cognitive architecture for multi-party conversation in an iCub robot; d) validate the real-time performance of the explainable model in multi-party human-robot interaction; e) propose several ways to incorporate explainability and transparency in the aforementioned architecture; and f) perform an online user study to analyze the effect of various explanations on how human participants perceive the robot.

Figures (17)

• Figure 1: Illustration of the addressee classification workflow. The sequence of faces is embedded using a vision transformer (M1), whereas poses are processed via an MLP. These embeddings are then fused using an intermediate network (M2), and their representation for each time frame is processed by a recurrent network enhanced with attention, forming a unified embedding of the whole utterance. The final step is the mapping to three output options via a fully-connected (FC) layer.
• Figure 2: The scheme of the recurrent network augmented with an attention mechanism. The input sequence (left) is projected to form keys, queries, and values. The queries are fed to the GRU network to obtain a single query. That is used to match the keys and compute the scores, which are used to sum up the values (forming the output).
• Figure 3: Attention maps extracted from the penultimate layer of the vision transformer employed in the model. The yellow areas indicate active information flow, whereas the blue areas correspond to patches that are not significant at the current layer.
• Figure 4: Visualization of two sequences, with their corresponding attention scores (dots) generated by the recurrent network, alongside the generated explanations. The dots' color and size correspond to the attention score magnitude for each time frame.
• Figure 5: scheme of the modular architecture implemented in the robot iCub. In pink are the audio processing modules, in light blue are visual processing, in yellow are modules to support reasoning processes, and in green are modules to act in the world.