[3N1-GS-8a] Robot and real worlds

10:00 AM - 10:15 AM JST(1:00 AM - 1:15 AM UTC)

[3N1-GS-8a-05] Joint Attention Control for a Listener Robot in Multi-Party Dialogues

〇Hibiki Baba¹, Yoshinari Onodera¹, Keijiro Onodera¹, Hanju Lee², Michita Imai¹ (1. Keio University, 2. DENSO CORPORATION)

Keywords:

Human-Robot Interaction (HRI),Multi-party dialogue,Joint attention,Nonverbal behavior,Social robotics

Conversational robots must show appropriate listener behavior in multi-party dialogue, not only in two-person interaction. Joint attention—shifting gaze toward a referent indicated by the speaker—signals empathy. In multi-party settings, candidate referents increase and gaze estimation can be unstable (e.g., occlusion), making behavior jittery. Simply following the speaker’s estimated gaze may turn estimation noise into frequent gaze switches, producing an unnatural impression. We propose Co-Attender, a joint-attention control system that balances social acceptability (attentive, natural behavior) and behavioral stability (calmness). Co-Attender focuses on speaker-led joint attention: it identifies the current speaker, estimates each participant’s gaze target with GazeLLE and face detection, and confirms the target using temporal gaze-following relations—whether a non-speaker later aligns to the same target. This reduces reliance on the speaker’s gaze estimate and suppresses transient fluctuations. In three-person dialogue, Co-Attender was robust to estimation noise and improved ratings of attentiveness/calmness over baselines (speaker-face fixation; speaker-gaze following).