Crosswalk Task: A naturalistic reinforcement learning paradigm for characterizing real-world risky behavior
The authors introduce the Crosswalk Task, a naturalistic reinforcement learning paradigm that utilizes inverse reinforcement learning to more effectively capture individual differences in real-world risk preference compared to conventional aggregate behavioral measures.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Every day, people make split-second choices that balance safety against speed. A pedestrian glancing at an oncoming car before stepping off the curb is weighing a potential reward—getting to the other side quickly—against a severe risk. Psychologists have long tried to understand why some people take these risks while others wait, but their traditional tools often fall short. In a laboratory, researchers might ask people to fill out a questionnaire about their willingness to take risks or have them play a simplified computer game with fixed rules. These methods, however, often fail to capture the messy, changing reality of the real world, where decisions depend on a shifting environment and immediate consequences. To bridge this gap between the lab and life, researchers needed a way to observe how people actually behave when the stakes feel real, without the artificial constraints of a standard experiment.
A team of researchers at Seoul National University and the Georgia Institute of Technology developed a new way to study this behavior called the Crosswalk Task. Instead of asking people what they would do, they created a digital simulation where participants guide a character across a series of busy roads. The goal is simple: reach the top of the map as fast as possible. But the path is filled with moving cars, and a collision ends the game immediately. The task forces players to make constant, dynamic choices. They can rush across the street, hoping the gap in traffic is wide enough, or they can wait for a designated crosswalk where the cars stop for them. Using the crosswalk is safer but takes more time, which might mean running out of the clock before reaching the finish line. Rushing is faster but carries the danger of a crash. By watching how people navigate these choices over and over again, the researchers could see their true risk-taking style in action.
To make sense of the thousands of movements and decisions made by 309 adults who played the game, the researchers used a sophisticated method called inverse reinforcement learning. In a standard computer program, you tell the machine what the reward is, and it learns how to get it. In this approach, the researchers did the opposite: they watched the humans play and worked backward to figure out what the humans were valuing. They built a model that inferred the invisible "reward" each person was chasing. Did this person value speed so much that they were willing to ignore the danger? Did another person value safety so highly that they wasted time waiting for a crosswalk that wasn't strictly necessary? This method allowed the team to uncover the hidden preferences driving the behavior, rather than just counting the final score.
The study found that these hidden preferences told a much clearer story than simple game scores did. When the researchers looked at the raw numbers—like how many times a player crashed or how high their total score was—the connection to a person's general risk-taking personality was weak. However, when they used the model to infer the specific rewards each person was seeking, the picture became sharp. The model revealed that people who scored high on self-reported risk-taking scales were indeed the ones who consistently chose to jaywalk, even when a safe crosswalk was nearby. More importantly, the model showed that individual differences in risk-taking propensity were more strongly reflected in observed road-crossing behavior under higher collision penalties. This nuance, where the link between personality and behavior became clearer as the stakes increased, was a detail that simple game scores missed entirely.
The researchers also discovered that combining different types of data gave the best results. They tested whether looking at just the game scores, just the inferred rewards, or a mix of both would better predict a person's risk personality. The mix of both behavioral data and the inferred rewards worked best, offering a more complete picture of individual differences than any single measure could provide. This suggests that to truly understand how people take risks, we need to look at both what they do and the invisible values driving those actions. The team made the software for this task freely available to other scientists, allowing them to study decision-making in a dynamic, naturalistic environment. By moving beyond static tests and into a world that changes and reacts, this work offers a new lens for seeing how we navigate the risks of everyday life.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.