Clustering Trust Dynamics in a Human-Robot Sequential Decision-Making Task

Cong Shi, Shreyas Bhat, Joseph B. Lyons, Cong Shi 0001, X. Jessie YangIEEE RA-L 2022 · DOI

TL;DR

This paper introduces a novel framework for human-robot trust-aware sequential decision-making using a Markov Decision Process (MDP) with a reward-based performance metric. They discovered three distinct human trust dynamics (Bayesian Decision Makers, Oscillators, and Disbelievers) and identified personal characteristics that correlate with these dynamics, paving the way for more personalized human-robot interactions.

What Problem Does This Paper Address?

Effective human-robot interaction hinges on trust, yet modeling and adapting to the dynamic nature of human trust remains a significant challenge. Previous models often relied on simplistic binary performance measures, which are insufficient for complex decision-making scenarios where cumulative rewards are key. Furthermore, while different trust dynamics have been observed, there's been a lack of research connecting these dynamics to individual human characteristics. This paper aims to bridge this gap by developing a more nuanced trust model and exploring the relationship between individual differences and how people's trust evolves over time during sequential decision-making tasks.

Methods

The core of the paper's methodology is a finite-horizon Markov Decision Process (MDP) framework designed for trust-aware sequential decision-making. A key innovation is the introduction of a reward-based performance metric to estimate the robot's performance. This metric acts as a surrogate for the robot's actual reward function, allowing the robot to infer human trust. The human's trust is modeled as a Beta distribution, with parameters updated in real-time using gradient descent based on the robot's performance feedback.

To understand how trust evolves, the researchers conducted a human-subject experiment where participants engaged in an Intelligence, Surveillance, and Reconnaissance (ISR) mission alongside an intelligent drone. After collecting trust data, they employed k-means clustering to group participants into distinct trust dynamic categories. Finally, they used one-way ANOVA to analyze the associations between these identified trust clusters and various personal characteristics measured through pre-experiment surveys.

Datasets

The paper does not explicitly mention using pre-existing datasets for training or evaluation. Instead, it introduces a newly collected dataset from a human-subject experiment involving 45 participants performing a simulated ISR mission. This dataset captures moment-to-moment trust ratings, task performance metrics, and individual personality traits.

Key Insights and Ideas

  • Reward-Based Trust is Key: The paper demonstrates that a robot's immediate reward can be a strong indicator of how a human's trust will change. This moves beyond simple success/failure metrics to a more nuanced understanding of trust dynamics.
  • Three Distinct Trust Personalities Emerge: The clustering analysis revealed three significant groups:
  • Bayesian Decision Makers: These individuals exhibit generally high trust and their trust is relatively predictable.
  • Disbelievers: Characterized by low trust and higher frustration, they are less likely to accept robot recommendations.
  • Oscillators: Their trust fluctuates rapidly, making it difficult to predict their behavior.
  • Personality Predicts Trust Dynamics: The study found significant correlations between personality traits and trust dynamics. For instance, Disbelievers tend to be less extroverted and agreeable, while Bayesian Decision Makers often have high expectations of automation.
  • Personalization is Crucial: The varying accuracies of the trust model across different clusters highlight the need for personalized trust models. A one-size-fits-all approach to trust in human-robot interaction is unlikely to be effective.

Key Results

  • The proposed reward-based performance metric was able to capture moment-to-moment trust changes with a prediction root mean squared error of 0.1266 (SD = 0.078).
  • K-means clustering successfully identified three distinct trust dynamics: Bayesian Decision Makers, Oscillators, and Disbelievers.
  • Significant differences were found in personality traits between the clusters, particularly in Extraversion (p = 0.011) and Agreeableness (p = 0.048).
  • Disbelievers were found to be significantly less extroverted than Oscillators (p = 0.009) and had significantly lower expectations of automation compared to both Bayesian Decision Makers (p = 0.005) and Oscillators (p = 0.023).
  • Post-experiment trust reports showed significant differences across all three clusters. Oscillators reported significantly higher frustration levels than Bayesian Decision Makers (p = 0.025).

Limitations and Future Work

Limitations:

  • The study acknowledges that the trust model achieved different accuracies for different trust dynamics, suggesting a need for tailored models for each category.
  • The experiment was conducted in a simulated environment, and real-world performance might differ.

Future Work:

  • The authors suggest exploring reverse-psychology models and other human trust-behavior models.
  • They propose using inverse reinforcement learning to learn human teammate preferences and improve team performance.
  • Further research could focus on developing personalized trust models that adapt to individual differences and specific trust dynamics.

Summary updated

BibTeX
@article{shi2022clustering,
  author = {Cong Shi and Shreyas Bhat and Joseph B. Lyons and Cong Shi and X. Jessie Yang},
  title = {{Clustering Trust Dynamics in a Human-Robot Sequential Decision-Making Task}},
  journal = {IEEE RA-L},
  year = {2022},
  doi = {10.1109/lra.2022.3188902},
}
Figure