A tale of two explanations: Enhancing human trust by explaining robot behavior.

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Explainability is crucial for AI and robot acceptance in critical applications.
• Current AI systems often lack comprehensive explanations for their actions.
• Human trust in intelligent systems is essential for widespread adoption.

Purpose of the Study:

• To investigate which explanation formats most effectively build human trust in AI and robot systems.
• To propose a framework for generating explanations from both functional and mechanistic viewpoints.
• To explore the relationship between explanation effectiveness and task performance.

Main Methods:

• A robot system learned to open medicine bottles from human demonstrations.
• Utilized an embodied haptic prediction model for sensory feedback knowledge extraction.
• Employed a stochastic grammar model for task compositional structure and an improved Earley parsing algorithm.
• Conducted a psychological experiment comparing different explanation types (visualizations vs. text summaries).

Main Results:

• The robot successfully learned and executed the bottle-opening task on unseen bottles.
• Real-time, comprehensive visualizations of the robot's internal decisions significantly enhanced human trust compared to text summaries.
• Explanation forms that best foster trust do not always align with components that optimize task performance.

Conclusions:

• Integrating diverse model components is necessary to simultaneously improve task execution and human trust in AI and robotics.
• Visual, real-time explanations are more effective for building trust than static text descriptions.
• A gap exists between optimizing task performance and optimizing human trust through explanations.