Active Inference Through Energy Minimization in Multimodal Affective Human-Robot Interaction.

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Humans use multimodal signals (e.g., facial expressions, gestures) to convey and interpret emotions.
• Robots need to efficiently process these signals for effective human-robot interaction, especially with limited resources.

Purpose of the Study:

• To propose an active perception method for robots to select the most informative emotional modalities.
• To enhance robots' ability to understand human emotional states using limited sensory input.

Main Methods:

• Developed an active perception method based on energy minimization.
• Utilized a multimodal deep belief network (an energy-based model) to link emotional states with sensory signals.
• Evaluated performance using criteria like information gain and active inference based on the free energy principle.

Main Results:

• The proposed method demonstrated improved accuracy in selecting informative modalities compared to other active perception approaches.
• The energy-based model effectively learned state probabilities, with lower energy indicating higher probability.
• Superior performance was observed in tasks involving mutually correlated multimodal information.

Conclusions:

• The active perception method enhances robots' ability to interpret human emotions by efficiently selecting key sensory modalities.
• This approach offers advantages in affective human-robot interaction, particularly when dealing with complex, correlated emotional cues.