PaperlayerView exact revision on arXiv

Source-linked AI summary

Do You Want Your Autonomous Car To Drive Like You?

Chandrayee Basu, Qian Yang, David Hungerman, Mukesh Singhal, Anca D. Dragan

arXiv:1802.01636v1cs.ROcs.HC

TL;DR

Autonomous cars may need to choose among many feasible driving behaviors, while common approaches assume they should match how users drive. This study tested that assumption in a simulator and found that users typically preferred a more defensive style than their own, often mistaking it for their own style.

  • Problem

    Many feasible driving behaviors can satisfy safety and feasibility constraints, yet a common approach assumes autonomous cars should match how their users drive.

    Source: §1 · p. 1

  • Method

    A simulator study compared participants’ own driving style with aggressive, defensive, and another user’s style, measuring preference and perceived similarity to their own.

    Source: §1 · pp. 1–2

  • Results

    Users preferred a more defensive style than their own, and over 80% preferred the style they thought was theirs despite often misidentifying their actual style.

    Source: §5 · p. 8

  • Takeaways & Limitations

    Learning users’ preferences may require more than driving demonstrations, since people can demonstrate how they drive rather than how they want to be driven.

    Source: §1 · p. 2, §5 · p. 8

    The study offered only a limited set of styles, so it identified the best among those options rather than the style users actually desired.

    Source: §5 · p. 8

Abstract

from arXiv · show

With progress in enabling autonomous cars to drive safely on the road, it is time to start asking how they should be driving. A common answer is that they should be adopting their users' driving style. This makes the assumption that users want their autonomous cars to drive like they drive - aggressive drivers want aggressive cars, defensive drivers want defensive cars. In this paper, we put that assumption to the test. We find that users tend to prefer a significantly more defensive driving style than their own. Interestingly, they prefer the style they think is their own, even though their actual driving style tends to be more aggressive. We also find that preferences do depend on the specific driving scenario, opening the door for new ways of learning driving style preference.

1. INTRODUCTION

The paper questions whether autonomous cars should reproduce how users drive, and tests whether users instead prefer a different style. In a simulator study, users generally preferred more defensive driving than their own, while preference tracked the style they thought they drove rather than their actual style.

1. INTRODUCTION

  • Autonomous cars may face many feasible ways to drive, reflecting individual trade-offs between efficiency and defensive behaviors such as keeping distance or braking quickly.The paper frames choosing a driving style as a question that follows the ability to generate safe, feasible behavior.Source: §1 · p. 1
  • The study challenges the assumption that autonomous cars should match users’ driving styles, an approach that prior work had pursued through Inverse Reinforcement Learning.The authors hypothesize that users may want a style different from their own.Source: §1 · p. 1
  • Participants tested their own style without knowing it, alongside aggressive, defensive, and another user’s style, after first demonstrating their driving in different environments.The study measured both preference and perceived similarity to their own style.Source: §1 · p. 2
  • Users typically preferred more defensive driving as passengers than their own measured style, while preferred style strongly correlated with the style they perceived as their own.There was little correlation between perceived and actual style.Source: §1 · p. 2
  • Learning how users actually want to be driven may require input beyond demonstrations, because users can demonstrate how they drive but may not demonstrate their preferred style.The authors also note that users may perceive a preferred style as their own, despite the mismatch with actual driving.Source: §1 · p. 2

2. RELATED WORK

Related work characterizes driving style in several ways, with defensiveness versus aggressiveness a common framing. Prior research on autonomous cars has explored both comfort and learning from demonstrations, generally assuming that cars should learn a user’s behavior.

2. RELATED WORK

  • Driving style is commonly framed along a defensiveness–aggressiveness dimension, although prior accounts also distinguish multiple styles and related behaviors.Examples include reckless, anxious, angry, and patient or careful driving, as well as violations, comfort, and lane-change characteristics.Source: §2 · p. 2
  • The paper aggregates driving features across scenarios to define defensiveness, extending beyond work focused on particular contexts such as lane changes, intersections, or car following.Features can include actions such as throttle and braking, and measures such as acceleration, deceleration, or maximum speed.Source: §2 · p. 2
  • Autonomous-driving research has examined how explicit styles relate to comfort and how cars can learn from human demonstrations; both lines generally assume the car should learn its user’s style or behavior..Source: §2 · p. 2

3. METHODS

The methods test whether autonomous-car users prefer a style different from their own. Participants first drove in a simulator, then experienced several styles whose defensiveness was defined and manipulated using driving features.

3.1 Hypothesis

  • The study tests the hypothesis that users prefer an autonomous car’s driving style to differ significantly from their own because being a passenger differs from being the driver.Source: §3.1 · p. 2

3.2 Study Design

  • The two-part simulator study first collected participants’ driving data, then used it to let them experience their own style in a later session.The study design was intended to test the hypothesis by having users experience and evaluate autonomous cars with different styles.Source: §3.2 · p. 2

3.3 Manipulated Variables

  • The manipulated styles were aggressive, defensive, a participant’s own, and another participant’s style, with participants unaware which style was theirs.Defensiveness was defined objectively from driving features; larger distances to other cars indicated more defensive driving, and the aggressive and defensive styles were created by demonstration and validated using those features.Source: §3.3 · pp. 2–3

3.4 Simulator and Driving Tasks

Participants drove a 14-task track and later rode as passengers while autonomous cars performed selected tasks in four styles. The procedure familiarized participants with the simulator before collecting their driving and style evaluations.

3.4 Simulator and Driving Tasks

  • The simulator used OpenDS with a classroom projection screen, table, steering wheel, brake, and gas pedal.Source: §3.4 · p. 3
  • Participants first drove a 9.6-mile track containing 14 tasks, each designed around everyday traffic conditions and sequences of maneuvers.Source: §3.4 · p. 3

3.5 Procedure

  • Before the first study part, participants practiced on two tracks until they felt they were driving as in everyday life, then completed the 14-task drive and an interview.Practice lasted 5–10 minutes; the drive lasted 15–20 minutes and was followed by a 10-minute interview.Source: §3.5 · p. 3
  • In the second part, autonomous cars performed six selected tasks combined into four test tasks, and participants completed an interview-based survey after experiencing each style on each task.Two test tasks were on expressways and lasted about four minutes per style; two inner-city tasks lasted under two minutes.Source: §3.5 · p. 3, §3.4 · p. 3

3.6 Dependent Measures

The study combined interviews and ratings of perceived driving similarity with task-specific driving features to assess driving style and preferences.

3.6 Dependent Measures

  • A post-driving interview asked whether simulated manual driving resembled daily driving, including a similarity rating from +3 to -3.Participants were also asked whether they enjoyed the drive and about positive or negative aspects of the simulator, controls, and traffic.Source: §3.6 · p. 3
  • Participants rated each driving style for comfort, safety, everyday-use preference, and similarity to their own on 7-point scales.They also thought aloud about their emotions and feelings while experiencing autonomous driving.Source: §3.6 · p. 3
  • The researchers used task-specific driving features drawn from prior work and participant interviews to quantify driving style.For Tasks 1 and 2, features included mean lead-car distance and headway measures; Task 1 also included merge-back distance headway.Source: §3.6 · p. 3
  • An aggregate Defensiveness Score averaged normalized feature values across each participant’s four test tasks.Scores were first computed for each participant and task.Source: §3.6 · p. 4

3.7 Manipulation Check

The manipulation check found that the aggressive and defensive reference styles differed in defensiveness, with most participants’ styles between them.

3.7 Manipulation Check

  • 86.67% of participants’ styles scored higher in defensiveness than the aggressive style and lower than the defensive style.The aggressive style had a lower defensiveness score than the defensive style.Source: §3.7 · p. 4

3.8 Participants

The study recruited 15 undergraduate and graduate students with varied driving experience and self-reported driving behaviors.

3.8 Participants

  • 15 participants, mostly aged 18–24, took part; three were 30–31 years old.The sample mixed graduate and undergraduate students and was screened for varied demographics, driving experience, and perceived driving behavior.Source: §3.8 · p. 4
  • Participants reported a mean of 5.46 years of driving experience (SD 4.5) and drove an average of 214 miles the preceding week (SD 188).Source: §3.8 · p. 4
  • The mean conservative–adventurous self-rating was 3.6, while the mean rating for adapting driving to environment and traffic was 3.Most participants placed themselves near the middle of the conservative–adventurous scale; only one called themself conservative.Source: §3.8 · p. 4
  • More participants favored comfort and steadiness over joy of driving, with a mean rating of 4.46.Source: §3.8 · p. 4

3.9 External Validity and Controlling for Confounds

The study used a simulator track designed to elicit natural driving and took steps to reduce recognition of participants’ own driving in the replay condition.

3.9 External Validity and Controlling for Confounds

  • The simulator track and traffic conditions were designed to elicit natural driving responses.The researchers used a simulator rather than real autonomous cars.Source: §3.9 · p. 4
  • To reduce recognition from familiar surroundings, replay retained the original route and traffic conditions but changed the scenes and traffic cars.Source: §3.9 · p. 5
  • Replay used only some of the manual-drive tasks, reordered tasks, and randomized style presentation to make participants’ own driving harder to recognize.Source: §3.9 · p. 5
  • After pilot participants recognized their driving because of simulator quirks, the researchers changed brake stiffness and steering sensitivity and smoothed replayed data.Source: §3.9 · p. 5

3.10 Trajectory Smoothing

The study filtered driving trajectories to remove instantly recognizable idiosyncrasies. It applied a Bilateral Filter to reduce lateral variance on expressway stretches while preserving distances between cars.

3.10 Trajectory Smoothing

  • The trajectories were filtered to eliminate idiosyncrasies that could make them instantly recognizable.Source: §3.10 · p. 5
  • A Bilateral Filter reduced lateral variance while preserving the distance between cars.Filtering was limited to expressway stretches.Source: §3.10 · p. 5
  • 15% smoothing reduced one participant’s trajectory lateral variance by 15%.Figure 4 compares the smoothed trajectory with the original for task 1.Source: §3.10 · p. 5

4. RESULTS

Participants generally rated the simulator as reasonably similar to real driving, and most participant styles fell between the aggressive and defensive reference styles. Overall, they preferred a more defensive style than their own, while preference also varied by task and aligned more with perceived than actual personal style.

4.1 Simulation Realism

  • Simulator ratings and comments supported conditions being not too far from real conditions; one participant objected to missing motion feedback, and participants with one or fewer years’ experience could not use simulation properly.The rating mode for similarity was +1 on a -3 to +3 scale.Source: §4.1 · pp. 5–6

4.2 Feature Distribution for Participant Styles

  • Most participants’ aggregate defensiveness scores fell between the reference styles, although some feature- and task-specific exceptions occurred.The feature plots arrange values from aggressive to defensive; participant marks are blue, aggressive-style values red, and defensive-style values green.Source: §4.2 · p. 6, Figure 3 · p. 6

4.3 Preferred Style in Relation to Own Style

  • Users preferred a significantly more defensive style than their own, with significant task effects indicating that preferences varied by context.The overall comparison was significant (t(1, 60) = −2.58, p = .0121); task effects were significant (F(3, 58) = 4.13, p = .0101).Source: §4.3 · pp. 6–7, Figure 5 · p. 7

in Relation to Actual Own Style

5. DISCUSSION

Users preferred more defensive driving than their own, and the discussion cautions against treating demonstrations as preferences. The authors identify limits in the tested driving styles, features, and simulation fidelity.

5. DISCUSSION

  • Users preferred a more defensive style than their own, echoing prior findings that people prefer lower speeds when not in control.Autonomous cars are one instance of not being in control of the driving.Source: §5 · p. 8
  • Over 80% of users preferred the style they thought was their own, although many misidentified their actual style.The authors caution that demonstrations reveal how people drive, not necessarily how they want to be driven.Source: §5 · p. 8
  • The study characterized driving style only by defensiveness and inherited its feature choices from prior studies.Source: §5 · p. 8
  • Participants chose among few defensiveness options, so the study found their best tested style rather than the style they actually desired.Source: §5 · p. 8
  • The simulator lacked motion feedback, which may limit speed perception; higher-fidelity experiments might yield more accurate results.Interviews validated that participants’ perceptions of the styles were sufficient.Source: §5 · p. 8
  • Future work could test broader features and style representations in higher-fidelity settings, examine the features users consider, and learn preferred-style deviations using additional user guidance.Source: §5 · p. 8
Loading 1802.01636v1…