Foundations for the psychological safety of human and autonomous vehicles interaction

Yandika Sirgabsou, Benjamin Hardin, François Leblanc, Efi Raili, Pericle Salvini, David Jackson, Marina Jirotka, Lars KunzearXiv 2024 · DOI · arXiv

Abstract

While current engineering standards for autonomous vehicles (AVs) focus primarily on physical safety and collision avoidance, the psychological well-being of users remains a significant barrier to widespread adoption. This paper introduces AV-PsySafe, a novel framework designed to bridge the gap between physical and psychological safety. By adapting established system-safety methodologies, the framework provides a systematic approach to identifying, quantifying, and mitigating psychological hazards—such as stress and loss of trust—during the vehicle design phase.

Introduction: The Adoption Gap

The "adoption gap" in autonomous driving is increasingly recognized as a psychological challenge rather than a purely technical one. Even when an AV is statistically safer than a human driver, passengers often experience vulnerability or fear. Current automotive safety standards, such as ISO 26262 and Safety of the Intended Functionality (SotIF), are engineered to prevent physical injury and property damage. However, they lack the formal tools to address "psychological hazards"—scenarios where a vehicle’s behavior is physically safe but mentally traumatic (e.g., an abrupt, unexplained swerve).

AV-PsySafe treats psychological safety as a formal engineering requirement, shifting the industry from reactive trauma measurement to proactive risk mitigation.

The AV-PsySafe Framework

The authors propose a three-part methodological framework to integrate psychological considerations into the system-theoretic safety process.

1. Formalizing Psychological Safety

The framework defines psychological safety in the context of AVs as the "absence of unacceptable psychological risk." Here, risk is calculated as a function of the probability and the severity of potential mental harm. This definition allows engineers to treat mental states with the same rigor applied to mechanical failures.

2. The Psychological Safety Integrity Level (PsySIL)

Inspired by the Automotive Safety Integrity Level (ASIL) found in ISO 26262, the authors introduce the Psychological Safety Integrity Level (PsySIL). This scale (categorized from A to D) allows for the classification of risk based on three key parameters:

  • Severity: The duration and intensity of the psychological effect (Short, Medium, or Long term).
  • Exposure: The frequency with which a user encounters the hazardous situation.
  • Controllability: The ease with which the human user or the system can mitigate the resulting stressor.

3. Adapted System-Theoretic Process Analysis (STPA)

To identify hazards, the framework adapts System-Theoretic Process Analysis (STPA), a top-down hazard analysis method. Rather than focusing on component failures, this adaptation identifies "Unsafe Control Actions" (UCAs) that negatively impact a driver’s mental state. For instance, a system providing a generic, non-descriptive error message during a malfunction is identified as a UCA because it degrades the user's mental model of the vehicle's state.

Illustrative Case Study: Highway Lane Change

To demonstrate the framework, the authors utilize a pedagogical scenario involving an SAE Level 4 AV. In this scenario, the vehicle experiences a braking malfunction during a highway lane change. While the ADS (Automated Driving System) successfully avoids a collision, it executes a sharp swerve and displays a generic error message to the passenger.

By walking through the four steps of the adapted STPA, the authors identified specific psychological losses, such as a permanent loss of trust in the system. In this specific highway example, the scenario was assigned a PsySIL A rating. Although the exposure was high, the severity was deemed moderate and the situation was considered controllable, resulting in a low-level but formal safety requirement for the system's communication interface.

Key Insights

  • Psychological Safety as an Emergent Property: Psychological safety is not a feature of a single component but an emergent property of the interaction between the human, the ADS controller, and the environment.
  • Explainability as a Safety Feature: A critical finding is that inadequate communication is an engineering failure. If a driver cannot understand why a vehicle is performing a specific maneuver, their mental model collapses, leading to psychological harm. Explainability, therefore, is a functional safety requirement.
  • Proactive Engineering: AV-PsySafe allows engineers to identify "loss scenarios" during the design phase, before a vehicle is ever deployed on public roads.

Limitations and Future Outlook

While AV-PsySafe provides a foundational structure, several challenges remain. The current framework assumes an "average" psychological baseline, yet individual variability—such as the difference between a seasoned test engineer and a first-time passenger—is significant. Furthermore, there is a potential for "safety trade-offs," where a maneuver required for physical safety (e.g., a violent swerve to avoid a pedestrian) may be inherently psychologically damaging.

Future work will focus on testing the framework against more complex urban driving conditions and exploring how these models can be generalized to broader fields of human-robot interaction (HRI).

Summary updated

BibTeX
@misc{sirgabsou2024foundations,
  author = {Yandika Sirgabsou and Benjamin Hardin and François Leblanc and Efi Raili and Pericle Salvini and David Jackson and Marina Jirotka and Lars Kunze},
  title = {{Foundations for the psychological safety of human and autonomous vehicles interaction}},
  year = {2024},
  eprint = {2411.05732},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.05732},
}
Figure