More Works from Our Lab

Reframing Social Presence in Human-Robot Interaction

A conceptual framing of social presence in HRI that this gaze modulation work builds on.

HRI 2025
Adaptive Gaze Modulation in Social Robots

The earlier reinforcement learning approach to attention regulation this paper extends with DGEI 2.0 and a unified GCRL policy.

IROS 2025
Real-Time Social Presence Modulation: An Audio-Centric Approach

Context assessment for social presence modulation through the audio channel, complementary to the gaze channel used here.

CASE 2025

A Goal-Conditioned Deep Reinforcement Learning Approach to Gaze Modulation for Attention Regulation

Nipuni H. Wijesinghe1, Maleen Jayasuriya1, David Hinwood2, Janie Busby Grant3, Damith Herath1
1Collaborative Robotics Laboratory, Faculty of Science and Technology, University of Canberra
2IA-Cobotics Lab, School of Engineering, RMIT University, Melbourne
3Faculty of Health, University of Canberra
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2026
Paper Video Google Scholar Cite
Mean gaze score over time for the dynamic GCRL policy compared against a random baseline, with the dynamic policy closely tracking a stepped reference trajectory of 0, 50 and 100.

One trained policy, three targets. Under the dynamic GCRL policy (blue), gaze engagement closely tracks a reference trajectory stepping through 0, 50 and 100, while a random, non-adaptive baseline (green) stays erratic throughout the same six-minute session.

Abstract

Effective human-robot interaction requires dynamic attention regulation: the ability to both capture and release human attention based on context, a capability current robotic systems largely lack. This paper presents a goal-conditioned reinforcement learning (GCRL) framework that enables a robot to adaptively modulate human gaze engagement, and consequently attention, across distinct levels through coordinated behavioural control. Unlike prior approaches that require separate models for each engagement level, this unified GCRL simultaneously learns to adjust multiple platform-specific behaviours, head movement, navigation, gestures and vocal parameters, on a SoftBank Pepper robot, based on real-time gaze feedback. Central to the approach is the Dynamic Gaze Engagement Index 2.0 (DGEI 2.0), which combines stationary gaze entropy with temporal consistency metrics to differentiate sustained engagement from scattered attention, providing fine-grained feedback for reward shaping. Controlled experiments with 30 participants demonstrate the framework's effectiveness: gaze engagement precisely tracked target levels of 0, 50 and 100, while subjective measures revealed significant differences across intensity conditions (p < 0.001, η² = 0.599 for attention and awareness). High-intensity behaviours achieved 90% attention agreement rates, while low-intensity modes suppressed visual engagement to near-zero levels. These results establish a validated framework for bidirectional attention regulation in HRI, a key enabler for robots that need to navigate when to engage versus when to fade into the background, essential for long-term human-robot coexistence.

Video Presentation

Results by Condition

In the between-subject study (n = 30, 10 per condition), each target level shaped gaze engagement and perceived interaction very differently. The same DGEI 2.0 metric and the same trained policy produced all three patterns below, only the goal changed.

Low (LGM) · release

Gaze score against time during the low gaze mode condition, showing a brief initial spike followed by suppression to near zero.

Visual engagement suppressed to near zero after a brief transition. The only questionnaire item with unanimous agreement was that Pepper's behaviour affected participants, pointing to unidirectional rather than reciprocal influence.

Stacked bar chart of questionnaire responses for the low gaze group.

Medium (MGM) · sustain

Gaze score against time during the medium gaze mode condition, showing sustained fluctuation around a moderate level.

Gaze fluctuated around a moderate, sustained level. 60% agreed they paid close attention while only 20% reported distraction, consistent with a robot that stays salient without overwhelming.

Stacked bar chart of questionnaire responses for the medium gaze group.

High (HGM) · capture

Gaze score against time during the high gaze mode condition, rising quickly and holding at an elevated level for the session.

Engagement rose quickly and stayed elevated. 90% paid close attention, none ignored Pepper, and over 80% agreed the robot attended to them in turn, the strongest bidirectional effect observed.

Stacked bar chart of questionnaire responses for the high gaze group.

One-way ANOVAs across the three conditions found significant differences on every measure: Attention and Awareness (F(2,27) = 20.203, p < .001, η² = 0.599), Behavioural Interdependence (F(2,27) = 55.862, p < .001, η² = 0.805), and Total NMSPI score (F(2,27) = 13.409, p < .001, η² = 0.498). Cronbach's alpha confirmed excellent reliability for both scales (α = 0.934 Attention/Awareness, α = 0.933 Interdependence). Attention and interdependence moved in opposite directions across conditions, highlighting that engagement and reciprocity are not the same thing to optimise for.

Read the Paper

BibTeX

@inproceedings{wijesinghe2026gcrlgaze,
  title     = {A Goal-Conditioned Deep Reinforcement Learning Approach to Gaze Modulation for Attention Regulation},
  author    = {Wijesinghe, Nipuni H. and Jayasuriya, Maleen and Hinwood, David and Busby Grant, Janie and Herath, Damith},
  booktitle = {2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026},
  organization = {IEEE/RSJ}
}