Toward Informed AV Decision-Making: Computational Model of Well-being and Trust in Mobility

Zahra Zahedi, Shashank Mehrotra, Teruhisa Misu, Kumar AkashIJCAI 2025 · DOI

TL;DR

This paper introduces a novel Dynamic Bayesian Network (DBN) model that infers the cognitive states of both autonomous vehicle (AV) users and other road users, specifically focusing on "well-being" and "trust." This model is then integrated into a causal inference framework to enable AVs to make decisions that balance user well-being, trust, and operational costs, moving towards more human-centered automation.

What Problem Does This Paper Address?

As autonomous vehicles (AVs) become more prevalent, their integration into our daily lives raises critical questions about how they interact with humans. Current AV development often overlooks the environmental and social implications, leading to public perception issues, as seen with the Paris e-scooter ban. This paper addresses the need for AVs to go beyond basic mobility and holistically consider the well-being and satisfaction of both their users and other road users. A key challenge is the dynamic and often unobservable nature of these cognitive states (like trust and well-being), which significantly influence human behavior and acceptance of AVs. Existing methods often rely on intrusive self-reports or physiological measures, which are impractical for real-time decision-making. This work aims to bridge this gap by developing a computational model that can infer these states and use them to inform AV decision-making.

Methods

The core of this paper is a Dynamic Bayesian Network (DBN) designed to infer latent cognitive states. The model captures the relationships between:

  • User's latent states: well-being ($w_k$), trust ($t_k$), and intention ($i_k$).
  • Other road user's latent state: well-being ($w_k^O$).
  • Observable actions: AV's action ($a_k^R$), other road user's action ($a_k^O$).
  • Action alignment: whether the AV's action aligns with the user's intention ($a_l^k$).

The DBN is structured into two components: an R-DBN (for when the AV is the contributor) and an O-DBN (for when the other road user is the contributor). These models are learned from observational data.

To enable informed decision-making, the DBN is extended into a Causal Inference Model (CIM) framework. This allows the AV to reason about the expected utility of different actions by defining utility functions that can maximize user well-being, user trust, or a trade-off between user well-being, other's well-being, and AV costs. The CIM uses the learned CPDs from the DBN to determine optimal policies.

Datasets

The paper describes an observational study conducted to collect data for training and evaluating the DBN models. This study involved participants interacting with a simulated self-driving scooter (AV) and delivery robots (other road users) in a web-based riding environment. The study design involved 2x2 factorial combinations of "yielding" and "unyielding" actions by both the scooter and the robot, resulting in 32 unique combinations across two rides. Data collected included self-reported measures of well-being and trust after each interaction event, as well as user intention before the scooter's action.

The paper does not mention using pre-existing datasets for training or evaluation.

Key Insights and Ideas

  • Holistic User Modeling is Crucial: The paper strongly emphasizes that effective AV integration requires understanding and responding to human cognitive states like well-being and trust, not just task-related performance.
  • Dynamic Inference is Key: Cognitive states are not static; they evolve with interactions. The DBN approach effectively models this dynamic nature.
  • Action Alignment Matters: The alignment between a user's intention and the AV's action significantly impacts user well-being and trust. This is a critical factor for AV decision-making.
  • Trade-offs are Inevitable: AV decision-making needs to balance multiple, sometimes conflicting, objectives: user well-being, user trust, other road users' well-being, and the AV's own operational costs.
  • Cost Sensitivity: The cost associated with AV actions (e.g., yielding vs. unyielding) plays a significant role in determining optimal policies, especially when costs are relatively low.

Key Results

  • The DBN model achieved an accuracy of 77% for inferring well-being, 67% for trust, and 95% for intention.
  • The model demonstrated that user well-being increases when the scooter's action aligns with the user's intention and decreases when it doesn't.
  • Trust also shows dynamic changes influenced by action alignment and accommodative actions.
  • Other's well-being increases when the scooter yields and decreases when it unyields.
  • For maximizing user well-being, the optimal policy suggests always yielding if no evidence is available. If evidence of previous well-being is available, the policy becomes conditional on the user's well-being level.
  • When maximizing user trust, the optimal policy is also to always yield without prior evidence, but aligning with user intention becomes important.
  • In trade-off scenarios, the optimal policy is sensitive to the cost difference between yielding and unyielding actions. For costs less than approximately 0.2, yielding is preferred; for higher costs, unyielding becomes more optimal, but this can negatively impact well-being.

Limitations and Future Work

  • Limited Interaction Events: The observational study involved only four interaction events per participant. The authors acknowledge this limitation, suggesting that the data may not fully capture long-term trust and well-being dynamics.
  • Dyadic Interactions: The study focused on dyadic interactions (scooter vs. robot) in sidewalk environments without formal rules. Future work should extend the model to handle interactions involving multiple road users and different mobility modes.
  • Longitudinal Data: Future research should explore longitudinal data collection to better model temporal dynamics of trust and well-being.
  • Domain-Specific Adaptation: The model's applicability could be broadened by incorporating more domain-specific variables and contextual information, and adapting input variables and data collection methods for different interaction domains.

Summary updated

BibTeX
@inproceedings{zahedi2025toward,
  author = {Zahra Zahedi and Shashank Mehrotra and Teruhisa Misu and Kumar Akash},
  title = {{Toward Informed AV Decision-Making: Computational Model of Well-being and Trust in Mobility}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/1149},
}

Loading related content...

Figure