Empirical game theory of pedestrian interaction for autonomous vehicles

Camara, Fanta; Romano, Richard; Markkula, Gustav; Madigan, Ruth; Merat, Natasha; Fox, Charles · 2018 · OpenAlex-citations

DOI: 10.5281/zenodo.3265804

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 programming autonomous vehicles (AVs) to interact effectively with human road users, specifically pedestrians. Current AV algorithms often prioritize absolute safety by yielding to any obstacle, which can lead to inefficient traffic flow as other users learn to exploit this behavior. The authors aim to develop a method for inferring human preference parameters—specifically regarding time delay and collision risk—from empirical data, allowing AVs to negotiate priority more realistically using game theory. The study expands on the "Sequential Chicken" model, a game-theoretic framework for two agents negotiating priority at an intersection. To measure human behavior, the researchers conducted a controlled experiment with 16 participants divided into eight pairs. Subjects played a board-game version of the model on a plus-maze grid, choosing speeds (1 or 2 squares per turn) to reach the intersection first while avoiding collisions. Two conditions were tested: a "natural" game focused on commuting efficiency and a "chocolate" game where explicit rewards (chocolates) were given to the winner, loser, or neither in case of a crash. The authors used Bayesian inference combined with Gaussian Process regression to fit the model’s parameters (crash utility and time value) to the observed player actions, accounting for human error through a noisy mixture model. The results indicated that the competitive "chocolate" game produced more collisions and reduced average time delays between players compared to the natural game. Gaussian Process analysis of the likelihood surfaces revealed that participants in both conditions exhibited a preference for time-saving over crash avoidance, with the effect being more pronounced in the chocolate game. The likelihood surfaces were roughly radial, suggesting that the ratio between crash utility and time value, rather than their absolute magnitudes, may be the critical factor in decision-making. The authors note that this counter-intuitive result likely stems from the artificial, game-like nature of the experiment, which encouraged participants to view the interaction as a competition rather than a safety-critical scenario. The significance of this work lies in providing a methodological framework for measuring and modeling human road-user behavior to inform AV controllers. By fitting game-theoretic models to empirical data, AVs can be programmed to predict human actions and negotiate interactions more optimally, avoiding the inefficiencies of purely defensive driving. The authors conclude that while the current experimental setup is simplified, the method can be extended to more realistic scenarios using real-world data, potentially allowing AVs to use credible threats of collision or other non-physical deterrents to maintain traffic flow while ensuring safety.

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.

• driver vru interaction

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Theoretical Contribution: computational model