Abstract
Interactive Task Learning (ITL) aims to enable robots to acquire new tasks through natural interactions with human instructors. However, a significant bottleneck exists: the "mental model gap," where humans lack an accurate understanding of a robot's internal knowledge, perceptions, and limitations. This research, presented at AAAI-21, explores transparency mechanisms and interaction frameworks designed to help human teachers diagnose robot failures and build more accurate mental models. By leveraging the Soar cognitive architecture and Collaborative Discourse Theory, the study demonstrates that providing robots with the means to reveal their internal states significantly improves the efficiency and intuitiveness of robot training for non-experts.
Introduction
In the field of robotics, making machines smarter is only half the battle. For robots to learn effectively from humans, the humans must first understand the robots. This "human-in-the-loop" challenge is central to Interactive Task Learning (ITL). When a human instructor attempts to teach a robot a new task—such as building a tower of blocks—they often project human-like understanding onto the machine. This leads to failures when the robot's internal representation of the world (e.g., its vocabulary or sensory limits) differs from the human's expectations.
The Mental Model Gap
A mental model is a person's internal representation of how a system works. In ITL, if an instructor does not know whether a robot perceives a "blue block" by its color or merely as a "small object," they cannot provide effective corrections when the robot fails. This research addresses how robots can actively bridge this gap by revealing their internal knowledge and intentions through transparency mechanisms.
Methodology
The research utilizes the Soar cognitive architecture, a symbolic system that allows for traceable and explainable decision-making. Key components of the approach include:
- Transparency Mechanisms: Implementation of question-answering (QA) and visualization tools within the "Rosie" robotic agent, allowing teachers to query the robot's perceptions in real-time.
- Collaborative Discourse Theory (CDT): Modeling interactions as a bidirectional dialogue rather than a sequence of one-way commands.
- Intentionality Taxonomy: A framework developed to categorize human intentions during teaching (e.g., defining a concept vs. evaluating competence).
- Human Subject Studies: Controlled experiments with non-experts to identify where mental models break down and how transparency tools can repair them.
Key Insights and Results
The study yielded several critical findings regarding human-robot interaction:
- The Vocabulary Trap: Instructors easily identify failures involving common terms (e.g., "I don't see the block"), but struggle when robots use internal or "hidden" terminology that doesn't align with natural language.
- Prediction as a Metric: The research posits that a human's ability to accurately predict a robot's next action is the most reliable metric for the quality of their mental model.
- Effectiveness of Transparency: Simple tools, such as a labeled camera feed or basic QA capabilities, significantly increased the accuracy of failure diagnosis among non-expert participants.
- Intent-Based Interaction: By recognizing why a human is asking a question, the robot can provide responses that specifically target and correct misconceptions.
Conclusion and Future Work
While current implementations rely on template-based inputs, future work aims to transition toward more flexible natural language interactions. The ultimate goal is to design robot responses optimized for improving human predictive accuracy, ensuring that the "dialogue of learning" leads to faster and more robust task acquisition. By making robots more transparent, we empower human teachers to become more effective mentors.
Summary updated