A Case Study of Trust on Autonomous Driving

Shili Sheng, Erfan Pakdamanian, Kyungtae Han, BaekGyu Kim, Prashant Tiwari, Inki Kim, Lu Feng

Abstract

As autonomous vehicles (AVs) have benefited society, understanding the dynamic change of humans' trust during human-autonomous vehicle interaction can help to improve the safety and performance of autonomous driving. We designed and conducted a human subjects study involving 19 participants. Each participant was asked to enter their trust level in a Likert scale in real-time during experiments on a driving simulator. We also collected physiological data (e.g., heart rate, pupil size) of participants as complementary indicators of trust. We used analysis of variance (ANOVA) and Signal Temporal Logic (STL) to analyze the experimental data. Our results show the influence of different factors (e.g., automation alarms, weather conditions) on trust, and the individual variability in human reaction time and trust change.

Introduction: The Problem of Dynamic Trust

Trust serves as the essential "glue" in human-automation interaction. In the context of autonomous driving, trust levels dictate how a driver utilizes the system: undertrust may lead to the rejection of safety-enhancing features, while overtrust can result in a dangerous lack of intervention during critical system failures.

A significant challenge in current research is that trust is not a static attribute but a dynamic, subjective mental state that fluctuates based on vehicle performance and environmental context. Traditional post-experiment surveys often fail to capture these temporal nuances, suffering from "recency bias" where participants primarily recall the final moments of a trial. This study addresses this gap by measuring trust in real-time and identifying the specific triggers—such as weather conditions or alarm failures—that drive these fluctuations.

Experimental Design and Methodology

To capture high-resolution behavioral data, the researchers utilized a Force Dynamics 401CR high-fidelity motion simulator. Nineteen participants (mean age ~22, 63% female) engaged in 16 distinct driving scenarios, each lasting 180 seconds.

The experimental setup was designed for multi-modal data acquisition:

  • Real-time Trust Feedback: Participants adjusted their trust level on a 1–5 Likert scale using buttons integrated directly into the steering wheel while driving.
  • Physiological Monitoring: To complement self-reported data, the team collected Galvanic Skin Response (GSR), heart rate variability (HRV), and eye-tracking data (pupil size).
  • Controlled Variables: The study manipulated three primary factors: alarm reliability (perfect, early false, late false, or missing), weather conditions (sunny vs. rainy), and driving mode (fully autonomous vs. semi-autonomous).
The experimental setup included a high-fidelity driving simulator integrated with physiological sensors (GSR, eye-tracking, PPG) and a custom GUI for real-time trust reporting.

Analysis and Findings

The researchers employed a dual analytical approach. Analysis of Variance (ANOVA) was used to identify group-level trends, while Signal Temporal Logic (STL)—a formal specification language—was applied to model the evolution of trust over time for individual participants.

The "Silent Failure" Penalty

The most significant finding was the impact of alarm reliability. ANOVA confirmed a significant main effect for alarm type ($p = 0.026$). Specifically, "missing alarms" (MMMM)—where a hazard exists but the system fails to alert the driver—damaged trust far more severely than false alarms. This suggests that human drivers are more forgiving of a "cry wolf" scenario than a system that fails to detect a blatant threat.

Box plot of trust levels across different alarm conditions. The 'MMMM' (Missing Alarms) condition shows the lowest median trust, highlighting the severe impact of silent system failures.

Environmental Stress vs. Reported Trust

Interestingly, while weather conditions did not significantly alter reported trust scores, they had a measurable impact on physiological markers. Pupil size and GSR peaks were notably sensitive to rainy conditions. This indicates that environmental stress is processed physiologically even if the driver does not consciously lower their trust rating, suggesting that "body logic" may provide a more nuanced view of driver state than self-reporting alone.

Individual Variability and Reaction Times

By applying STL and the TeLEx tool, the researchers identified an "individual signature" in how trust evolves. The reaction time for trust adjustments varied significantly across the cohort, ranging from 1.0 second to 9.5 seconds. STL analysis also revealed that trust is most likely to increase after approximately 30 seconds of error-free autonomous driving, indicating that reliability over time acts as a slow "recharge" for human trust.

Conclusion

This study demonstrates that trust in autonomous systems is a complex, time-varying signal influenced by both system performance and environmental factors. While missing alarms are the primary "trust-killers," individual differences in reaction time and the disconnect between physiological stress and reported trust highlight the need for more personalized trust modeling. Future work aims to refine the STL framework to predict trust drops using physiological signals in real-time, potentially allowing systems to intervene before a driver consciously loses confidence.

Summary updated

Abstract

As autonomous vehicles have benefited the society, understanding the dynamic change of humans' trust during human-autonomous vehicle interaction can help to improve the safety and performance of autonomous driving. We designed and conducted a human subjects study involving 19 participants. Each participant was asked to enter their trust level in a Likert scale in real-time during experiments on a driving simulator. We also collected physiological data (e.g., heart rate, pupil size) of participants as complementary indicators of trust. We used analysis of variance (ANOVA) and Signal Temporal Logic (STL) to analyze the experimental data. Our results show the influence of different factors (e.g., automation alarms, weather conditions) on trust, and the individual variability in human reaction time and trust change.
BibTeX
@inproceedings{sheng2019case,
  author = {Shili Sheng and Erfan Pakdamanian and Kyungtae Han and BaekGyu Kim and Prashant Tiwari and Inki Kim and Lu Feng},
  title = {{A Case Study of Trust on Autonomous Driving}},
  booktitle = {ITSC},
  year = {2019},
  url = {https://openreview.net/forum?id=ZO8Gvu2r2eH},
}
Figure