Social Cohesion in Autonomous Driving

Landolfi, Nicholas C.; Dragan, Anca D. · 2018 · Crossref

DOI: 10.1109/iros.2018.8593682

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of autonomous vehicles operating in complex, real-world driving scenarios where they may lack specific knowledge of local rules, hidden obstacles, or unwritten traffic norms. The authors argue that while autonomous cars may fail to detect certain hazards or adapt to local driving cultures, human drivers around them often navigate these situations correctly. The research proposes a "socially cohesive" driving algorithm that leverages the consistent behavior of nearby human drivers to improve the safety and social acceptability of autonomous vehicles. The core intuition is that if human drivers consistently exhibit a specific behavior, the autonomous car should likely adopt that behavior as well, even if it cannot explicitly identify the cause. The method involves augmenting the autonomous vehicle’s reward function with a social cohesion term. The authors define social features based on the trajectories of other vehicles, such as changes in position and orientation. They group these features based on spatial proximity (e.g., cars in the same lane or nearby on the road) and calculate the variance of these features within each group. The autonomous car uses Model Predictive Control to optimize a reward function that balances its original driving objectives (avoiding collisions, staying in lane) with a cohesion term. This term incentivizes the car to match the mean behavior of a group only when the variance of that behavior is low, indicating strong agreement among human drivers. The system dynamically decides what to follow, who to follow, and when to follow based on this variance metric. The authors evaluate the algorithm through simulations across four diverse scenarios: avoiding a stalled car the robot cannot see, yielding to an ambulance, adapting to local speeding norms, and navigating a highway exit. In the first three scenarios, the cohesive car successfully mimics human behavior to avoid collisions or adapt to local norms, whereas a baseline car fails to yield or crashes. However, in the highway exit scenario, the cohesive car is tricked into taking an exit because all cars in its lane did so, demonstrating a potential false positive. A user study involving 12 participants assessed attitudes toward the cohesive car versus the baseline car. Results showed that participants significantly preferred the cohesive car, reporting higher perceptions of safety, social belonging, and overall passenger experience, while reporting lower frustration. Participants were also more willing to share the road with the cohesive car and judged its passengers less harshly. Notably, users remained tolerant of the cohesive car’s mistakes, such as taking the wrong exit, in exchange for the benefits of safer and more socially acceptable driving in other scenarios. The significance of this work lies in demonstrating that social cohesion can serve as a robust, generalizable mechanism for improving autonomous driving performance in the near term. By leveraging implicit information from human drivers, autonomous vehicles can overcome limitations in perception and rule knowledge, potentially accelerating their deployment. The findings suggest that humans are willing to tolerate occasional errors from socially cohesive cars if the overall driving experience is safer and more aligned with social norms. This approach offers a practical solution for bridging the gap between current autonomous capabilities and the complex, unwritten rules of human traffic systems.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• automation surprise