Personalizing Autonomous Driving:

How Adaptive AI Tailors Driving Style for an Individualized Experience

By Mariah Schrum, Andrew Best, and Emily Sumner

Introduction

The way we drive is often a reflection of our personalities. It might surprise you to learn that research has shown that our driving style is closely linked to who we are as individuals [1]. Adventurous types prefer the thrills and challenges of driving, while others prefer a more cautious approach. And, because people drive differently, it’s only natural they would also prefer different driving styles for autonomous vehicles (AVs).

This diversity presents both a challenge and an exciting opportunity for AV technology. One-size-fits-all models that produce the same AV driving style will likely not be ideal. But what if an AV could adapt and adjust its driving to suit the specific preferences of its user? This is where MAVERIC (Manipulating Autonomous Vehicle Embedding Region for Individual Comfort) comes in. At Toyota Research Institute (TRI), we developed MAVERIC, a data-driven approach that learns a person’s driving style from their data, allowing the autonomous vehicle (AV) to mimic their driving behavior. MAVERIC includes a “knob” that lets us easily adjust how assertive the driving is, offering a tailored experience that balances user style with other relevant factors that may impact preference.

At its core, MAVERIC is about personalization. While previous works have introduced methods to imitate an end user’s driving style, we aim to go beyond simply copying an end user’s style. Research shows that while some users might want their AV to drive exactly like they do, others prefer a style that differs from theirs. At TRI, we hypothesize that the optimal driving style for an AV is a function of both the end-user’s own driving style and various subjective factors, such as personality. We express this relationship in a simple formula.

𝑤ᵖ*(optimal driving style) = 𝑤ᵖ (end-user driving style) + f(c) (influence of subjective factors)

1. Develop a method to characterize the driving style (𝑤ᵖ) of an end-user via a data-driven approach.
2. Enable an AV to leverage this representation to both mimic and modulate driving style.
3. Investigate the factors that influence the relationship between the end user’s own driving style (𝑤ᵖ) and their preferred AV driving style (𝑤ᵖ*).

MAVERIC Framework

Characterizing Driving Style: Traditionally, driving style has been characterized by metrics developed in previous research (e.g., high average velocity and frequent lane changes = aggressive style) [2, 3]. However, our goal is to let a representation of driving style naturally emerge from driving data rather than characterizing it based on our preconceived notions. To achieve an unsupervised representation of driving style, MAVERIC learns a personalized driving style embedding (i.e., a vector representation of driving style) from end-user driving data, guided only by a questionnaire about the end user’s perception of their own style. With a representation of the end-user’s driving style, the AV can use this information to mimic their driving behavior accurately.

Mimicking Driving Style: The backbone of MAVERIC is an imitation learning framework (Fig. 1) conditioned on the learned personalized embedding (𝑤ᵖ). MAVERIC is trained to produce high-level behaviors so that the AV mimics the driving style of an end-user. These behaviors are then fed into low-level controllers that execute the behaviors. By separating low-level control and high-level decision-making, we improve training stability and make it possible to impose safety constraints. While accurately mimicking a user’s driving style is crucial, it’s just the first step — next, we explore how to adjust and fine-tune these behaviors to meet individual preferences, especially when users may want something different from their style.

Fig. 1. Overview of MAVERIC framework. We collect data from a set of participants to learn the model parameters and participants’ personalized driving embeddings. Then, at test time, we freeze these model parameters and learn only the personalized embedding for the new participant. This personalized embedding is then fed into the MAVERIC architecture to enable the AV to mimic the driving style of the participant. In other words, at test time, our goal is to determine where an end-user falls within the population of drivers and leverage this information to inform AV driving style.

Modulating Driving Style: While MAVERIC is trained to mimic the driving style of an end-user, research indicates that some end-users may prefer driving styles that differ from their own [4]. Driving style is multi-dimensional and can be modulated in many ways. We focus on modulating assertiveness as research has shown that the level of assertiveness significantly impacts end-user preference [3]. MAVERIC allows us to fine-tune the AV’s assertiveness by moving along a gradient of assertiveness within the embedding space. Want a bolder, more assertive style? We can dial assertiveness up by moving in the positive direction of the assertiveness gradient (Fig 2). Prefer a more laid-back, cautious ride? We can dial assertiveness down. With the ability to modulate assertiveness, the next question is: what subjective factors can help us predict whether someone will prefer a more or less assertive AV driving style compared to their own?

Fig 2. Embedding space. The line represents the gradient of assertiveness, and each point represents an end-user’s unique driving style. We show that the gradient of assertiveness is correlated with average speed, a metric often associated with assertive driving [5].

Factors Impacting Preference

With a framework in place that can both mimic and modulate driving styles, our next goal is to identify the factors that shape end-user preferences — specifically, determining the variable 𝑐 in Eq 1 that explains the difference between one’s preferred AV driving style (𝑤ᵖ*) and one’s own driving style (𝑤ᵖ). To do this, we measure 𝑤ᵖ* by presenting end-users with driving styles that differ from theirs and asking them about their preferences. We derive 𝑤ᵖ from their actual driving data, as discussed earlier.

We uncovered several key factors in a study involving more than 50 participants in TRI’s high-fidelity driving simulator (featuring a 6-DOF platform based on CARLA, ROS2, and Unreal Engine). The findings reveal that personality traits like conscientiousness and openness, perceived similarity, and users’ own perceptions of their driving style are key predictors (Fig 3).

Fig 3. Factors influencing driving style preference.

Our work provides critical components for determining the optimal driving experience for an individual end-user. Given these variables, we’re better equipped to fine-tune AV driving behavior to meet users’ unique preferences. Moving forward, our research will refine this relationship and explore how we can further enhance personalization in autonomous driving.

For a more detailed overview of MAVERIC, please see our paper: https://ieeexplore.ieee.org/abstract/document/10415518

[1] Wang Y, Qu W, Ge Y, Sun X, Zhang K. Effect of personality traits on driving style: Psychometric adaption of the multidimensional driving style inventory in a Chinese sample. PLoS One. 2018 Sep 6;13(9).

[2] Suzanne E Lee, Erik C B Olsen, Walter W Wierwille, and Michael Goodman. A comprehensive examination of naturalistic lane-changes, 2004.

[3] Chandrayee Basu, Qian Yang, David Hungerman, Mukesh Sinahal, and Anca D. Draqan. Do you want your autonomous car to drive like you? In 2017 12th ACM/IEEE International Conference on Human-Robot Interaction, pages 417–425, 2017.

[4] Fredrick Ekman, Mikael Johansson, Lars-Ola Bligård, Mari Anne Karlsson, and Helena Strömberg. Exploring automated vehicle driving styles as a source of trust information. Transportation Research Part F: Traffic Psychology and Behaviour, 2019.

[5] Zhizhuo Su, Roger Woodman, Joseph Smyth, and Mark Elliott. The relationship between aggressive driving and driver performance: A systematic review with meta-analysis. Accident Analysis & Prevention, Volume 183. 2023.