AgeLab Researchers Give Presentations on Driver Self-Regulation and Qualitative LLM Analysis at TRB

AgeLab researchers Pnina Gershon, Lisa D’Ambrosio, and Zhouqiao Zhao presented at the Transportation Research Board (TRB) 2026 Annual Meeting in Washington, D.C., in January. Part of the National Academies of Sciences, Engineering, and Medicine, the TRB convenes professionals to solve a variety of transportation-related challenges.

Gershon organized and moderated a session at the conference titled “Driver Competency Across the Lifespan: Assessment, Technology, and Policy” alongside John Lenneman, a research scientist at Toyota. The session focused on understanding drivers’ competency as dynamic and influenced by a range of factors like their age, health, or environment, as well as the specific vehicle they are driving. The session delivered a broad perspective by bringing together experts from different domains, including Jon Wey of State Farm Insurance, Bayliss Camp of the California Department of Motor Vehicles, Alexandra Muller of the Insurance Institute for Highway Safety, and Lisa D’Ambrosio from the AgeLab.

D’Ambrosio began her panel presentation, titled “Health, Self-Regulation & Driving,” by highlighting the importance of driving to people’s overall mobility, as well as their sense of autonomy and independence. There are just under 53 million drivers age 65 or older in the US, representing 22% of all drivers. This number is only expected to increase throughout the rest of the decade, raising the importance of designing systems to keep the older population mobile as they experience changes to their health that can affect their ability to get around safely, including by driving.

One behavior that can emerge as driving competency changes over a person’s life is self-regulation. Self-regulation is when a person reduces when or where they drive for safety reasons. For instance, they may not be willing to drive in the dark, on busy roads, or in places that require difficult maneuvers like merges or left turns. Self-regulation generally tends to work, making it a worthwhile strategy to adopt to manage health declines or other functional changes that affect driving.

D’Ambrosio continued by showing data from an MIT AgeLab study comparing levels of driving self-regulation across groups of different demographic characteristics. Women tended to self-regulate more overall than men did. The youngest group of drivers in the study—who may have had less competence due to less experience—self-regulated the most. Drivers who said they were in fair or poor health tended to self-regulate more than healthier drivers.

Each day, more vehicles that can handle an increasing number of driving tasks autonomously appear on the road. These new vehicles have the potential to change how people evaluate and regulate their own driving, raising the importance of understanding how driving competency and self-regulation vary throughout people’s lives.

Zhao presented a joint paper with Gershon titled, “Chain-of-Thought Multimodal Large Language Model for Driver–Pedestrian Negotiation.” This talk was given during the Chinese Overseas Transportation Association (COTA) Winter Symposium as part of the Young Professionals Lightning Talk. It focused on the potential for multimodal large language models (LLMs) to help researchers better understand situations on the road where drivers and pedestrians cross paths, such as when a driver approaches a crosswalk and a pedestrian steps toward the curb. Interactions like these are complex and unfold according to a combination of cues like how fast a car moves and in what direction, or how a driver and pedestrian look at each other. Because of this complexity, multimodal LLM applications for modeling driver–pedestrian interactions remain relatively limited.

Using driving scene data from the MIT Advanced Vehicle Technology (AVT) dataset, Zhao displayed a fine-tuned multimodal LLM’s capacity to identify the target pedestrian and other relevant road users, summarize a scene’s key context, predict the outcome of driver-pedestrian interactions, and produce chain-of-thought explanations for their predictions that reference observable cues like the ones mentioned above. Zhao’s demonstration is pioneering, expanding the range of emerging tools researchers can use to learn more about driver–pedestrian interactions.