DIARC: A Testbed for Natural Human-Robot Interaction

Paul W. Schermerhorn, James F. Kramer, Christopher Middendorff, Matthias ScheutzAAAI 2006

Abstract

The Distributed Integrated Affect, Reflection, Cognition (DIARC) architecture is designed to bridge the gap between low-level robotic control and high-level social interaction. DIARC addresses two primary challenges in Human-Robot Interaction (HRI): the complexity of multi-modal communication and the necessity for system robustness in unpredictable environments. By utilizing a reflective middleware within the Agent Development Environment (ADE), DIARC enables robots to detect and recover from component failures autonomously. Furthermore, the integration of bimodal emotion recognition and an affective action interpreter allows the robot to prioritize tasks based on human emotional states and its own performance history. This paper details the technical components of DIARC and its performance in the 2006 AAAI Robot Competition.

1. Introduction

Developing robots capable of natural interaction with humans remains a significant challenge in robotics. Effective HRI requires more than just speech recognition; it necessitates a seamless integration of navigation, object recognition, and the ability to interpret non-verbal social cues. The authors identify two critical requirements for modern HRI systems:

  1. Interaction Complexity (R1): The ability to process natural language, dialogue structures, and affective states (emotions) simultaneously.
  2. System Robustness (R2): The capacity to maintain social engagement even when underlying hardware or software components fail.

DIARC was developed as an integrated architecture to meet these requirements, providing a framework where cognitive processes and system-level reflection work in tandem.

2. The DIARC Architecture and ADE

DIARC is built upon the Agent Development Environment (ADE), a Java-based infrastructure that provides the distributed computing power necessary for complex robotic tasks.

2.1 Reflective Middleware

A core innovation of DIARC is its "reflective" middleware. Unlike traditional systems that may crash upon a module failure, ADE treats the system's own health as a set of logical facts. A reasoning component monitors these facts; if a failure is detected (e.g., a speech recognition module stops responding), the system applies predefined policies to migrate the task to a different processor or adjust the robot's behavior. This ensures that the robot remains functional and socially present even during technical glitches.

2.2 Affective Action Interpreter

Decision-making in DIARC is handled by an affective action interpreter. Rather than relying on a static decision tree, the robot maintains a prioritized goal stack. The utility or importance of these goals is dynamic, influenced by the robot's "affective state"—a representation of its recent successes and failures. This allows the robot to prioritize critical tasks or simplify its behavior when it encounters difficulties, effectively using "mood" as a computational filter to manage complexity.

3. Perception and Emotion Recognition

To interact naturally, DIARC employs several advanced perceptual modules to close the loop between human state and robot action.

3.1 Bimodal Emotion Detection

The system monitors human stress through two primary channels:

  • Acoustic Analysis: The robot tracks changes in voice pitch and intensity relative to an established speaker baseline established during an initial training phase.
  • Visual Analysis: Using a swarm-based search algorithm, the system tracks facial features, specifically the eyebrows and mouth, to identify expressions of surprise, anger, or frustration.

3.2 Perceptual Learning

For spatial awareness and object recognition, DIARC utilizes Scale Invariant Feature Transform (SIFT). This allows the robot to extract distinctive features from its environment, enabling it to learn new objects on the fly and respond to complex spatial queries, such as "Is the blue box to the left of the red one?"

4. Evaluation: The 2006 AAAI Robot Competition

The DIARC architecture was evaluated during the 2006 AAAI Robot Competition. The robot participated in a simulated "extraplanetary collaborative exploration" task, which required it to navigate a complex environment, identify objects, and follow natural language instructions from a human teammate.

4.1 Results

The system demonstrated high qualitative success. It successfully followed directions in a search-and-rescue context and adjusted its level of autonomy based on the detected stress levels of the human partner. If the human appeared overwhelmed, the robot would offer more proactive suggestions or take over specific sub-tasks. Crucially, the ADE infrastructure allowed the robot to survive component crashes that would have typically ended a mission, proving the efficacy of the reflective middleware in maintaining mission continuity.

5. Discussion and Future Work

DIARC represents a significant step toward resilient HRI. By treating fault tolerance as a cognitive problem and using affect to manage goal prioritization, the system achieves a level of robustness and naturalism rare in integrated systems of its time.

However, several limitations remain. The current bimodal emotion detection is a baseline for social interaction but lacks the nuance required for truly complex human social dynamics. Additionally, the system's performance is heavily dependent on parallel hardware, which posed a bottleneck for the embedded systems available during its initial development. Future work will focus on expanding the range of social cues the robot can interpret and improving the scalability of the affective reasoning components across more diverse environments.

Summary updated

BibTeX
@inproceedings{schermerhorn2006diarc,
  author = {Paul W. Schermerhorn and James F. Kramer and Christopher Middendorff and Matthias Scheutz},
  title = {{DIARC: A Testbed for Natural Human-Robot Interaction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure