Algorithmic Accuracy as a Motivational Driver in Robot-Mediated Learning: A Comparative Study of Cross-Correlation and CNN-Based Sound Detection in an Interactive Quiz Game

Technical Summary: Algorithmic Accuracy as a Motivational Driver in Robot-Mediated Learning

Problem Statement
While educational Human-Robot Interaction (HRI) research has extensively explored how robot appearance, behavior, and pedagogical strategies influence learner motivation, there is a significant gap in understanding the impact of the robot's underlying perception algorithms. In competitive learning scenarios, such as robot-mediated quiz games, inaccurate perception (e.g., failing to detect the first student to press a buzzer) can lead to unfair interactions, reduced trust, and diminished learner confidence. Existing sound classification research focuses on engineering metrics (accuracy, latency) in controlled environments, often neglecting how algorithmic performance degradation in real-world acoustic conditions translates to learner experience. Specifically, deep learning approaches like Convolutional Neural Networks (CNNs) often suffer from domain mismatch when deployed in acoustically different environments compared to their training data, whereas template-based methods may offer greater robustness without extensive retraining.

Methodology
The study employed a controlled between-subjects experimental design to investigate the relationship between sound detection accuracy and student motivation.

• Participants: 40 university students were equally divided into two groups (n=20 per group).
• System: A Pepper humanoid robot acted as the quiz host, managing question presentation, scorekeeping, and feedback. The system included an audio acquisition module and a sound detection module.
• Experimental Conditions:
  • Control Group (CNN): Utilized a Convolutional Neural Network trained on laboratory data without additional retraining or calibration for the specific classroom environment.
  • Experimental Group (Cross-Correlation): Utilized a Cross-Correlation algorithm that compared incoming audio signals against a reference template recorded immediately before each session, allowing adaptation to the specific acoustic environment without retraining.
• Procedure: Both groups participated in identical interactive buzzer-based quiz sessions. The only variable was the sound detection algorithm used to identify the first responder.
• Metrics:
  • Motivation: Assessed using the Intrinsic Motivation Inventory (IMI), measuring five subscales: Interest/Enjoyment, Perceived Competence, Effort/Importance, Perceived Choice, and Pressure/Tension (reverse-coded).
  • Algorithm Performance: Evaluated via real-time detection accuracy, latency, and the need for calibration.

Key Contributions

1. Comparative Evaluation: A direct comparison of CNN and Cross-Correlation algorithms for real-time first-responder identification in an educational setting, highlighting the trade-offs between data-driven learning and template-based signal processing.
2. Empirical Investigation: An analysis of the correlation between algorithmic accuracy and intrinsic motivation, demonstrating that technical reliability directly influences psychological outcomes.
3. APMR Model Validation: Empirical support for the proposed Algorithmic Precision–Motivation Relationship (APMR) model, which posits that improvements in perception accuracy enhance learner motivation by fostering perceptions of fairness, competence, and engagement.

Results

• Algorithm Performance: The Cross-Correlation algorithm demonstrated superior robustness in the classroom environment, achieving an 87% detection accuracy compared to 69% for the CNN. The Cross-Correlation approach required only a short template recording for calibration, whereas the CNN required prior training and showed no adaptation to the new acoustic environment. Latency was comparable (280 ms vs. 300 ms).
• Motivational Outcomes: Participants in the Cross-Correlation group reported significantly higher scores across all IMI subscales compared to the CNN group:
  • Interest/Enjoyment: Increased from 3.49 (CNN) to 4.38 (Cross-Correlation).
  • Perceived Competence: Increased from 3.74 to 4.31.
  • Effort/Importance: Increased from 3.80 to 4.60.
  • Perceived Choice: Increased from 4.13 to 4.27.
  • Pressure/Tension (Reverse-coded): Improved from 3.86 to 4.34, indicating lower psychological pressure.
• Qualitative Observation: The higher accuracy of the Cross-Correlation system resulted in fewer interruptions and misclassifications, leading students to perceive the robot's decisions as more consistent and fair.

Significance and Claims
The paper claims that algorithmic accuracy is not merely an engineering performance metric but a critical design variable in educational robotics. The findings suggest that the reliability of perception algorithms is as important as social behaviors in maintaining learner engagement and motivation. By validating the APMR model, the study argues that future educational robotic systems must prioritize robust perception mechanisms—such as environment-adaptive template matching over static deep learning models in variable acoustic settings—to ensure fair and motivating learning experiences. The authors note limitations regarding the sample size and the specific context of a single robot platform, suggesting future work should explore diverse populations, additional modalities (e.g., vision), and advanced deep learning architectures with domain adaptation.