Pose Generation for Social Robots in Conversational Group Formations.

Area of Science:

• Robotics
• Human-Robot Interaction
• Artificial Intelligence

Background:

• Robots participating in social human conversations require appropriate poses.
• Environmental layout significantly influences robot pose selection in group formations.

Purpose of the Study:

• To develop and evaluate two distinct methods for predicting robot poses in social conversational settings.
• To assess the effectiveness of model-based and data-driven approaches in handling environmental constraints and capturing formation dynamics.

Main Methods:

• A model-based approach explicitly encoding geometric aspects of conversational formations.
• A data-driven approach using graph neural networks and adversarial training for implicit spatial arrangement modeling.
• Evaluation through quantitative metrics and a human experiment.

Main Results:

• Both methods effectively reason about environment layout and conversational group formations.
• The geometric approach excelled at avoiding non-free space, while the data-driven method better captured formation variability.
• Methods can be repeatedly used to simulate spatial arrangements, despite single-pose output design.

Conclusions:

• The proposed methods offer effective solutions for robot pose prediction in social contexts.
• Different approaches have complementary strengths, suggesting potential for hybrid models.
• Further research is needed to address open challenges in robot pose generation for dynamic social environments.