Master’s Thesis

Adaptive In-Car Interfaces

Adaptive In-Car Interfaces

Should a car’s touchscreen suggest your next action, or just do it for you?

Should a car’s touchscreen suggest your next action, or just do it for you?

Advisory version: Suggested Actions panel with Apply button over the map

6 months

Timeline

Sole researcher and designer

Role

M.Sc. User Experience Design TH Ingolstadt

Context

Driving simulator study

Methods

Figma, JASP, MAXQDA

Tools

Prof. Andreas Riener

Supervisor

6 months

Timeline

Sole researcher and designer

Role

M.Sc. User Experience Design TH Ingolstadt

Context

Driving simulator study

Methods

Figma, JASP, MAXQDA

Tools

Prof. Andreas Riener

Supervisor

Thesis title: Design and Evaluation of Context-Adaptive Suggested Actions in In-Vehicle Interfaces

Introduction

Introduction

The version with the fewest taps was not the easiest one. Drivers wanted help with the action, and they wanted to keep the decision.

Car touchscreens now hold navigation, media, climate and vehicle settings in one place. Every tap takes attention that the road also needs. One answer is an interface that reads the driving context and offers the right action at the right moment: a school route suggests the child lock, a long trip suggests your road trip playlist.

Car touchscreens now hold navigation, media, climate and vehicle settings in one place. Every tap takes attention that the road also needs. One answer is an interface that reads the driving context and offers the right action at the right moment: a school route suggests the child lock, a long trip suggests your road trip playlist.

That raises a design question research had not settled. How much should the system do on its own? Most studies of adaptive car interfaces measure glance time, lane keeping and task time. Far fewer ask how it feels to the driver when the interface takes the initiative.

At a glance

12

12

drivers, each tested all three versions while driving

60

60

screens prototyped in Figma across 12 task flows

81

81

coded interview segments, grouped into 7 themes

0 of 12

0 of 12

named the fully automatic version as their pick for a real car

Project Overview

Project Overview

1

The goal

Cut the taps needed for common in-car tasks without making the interface feel unpredictable.

1

The goal

Cut the taps needed for common in-car tasks without making the interface feel unpredictable.

2

The challenge

Automation saves effort but takes decisions away. In a car, a surprise on the screen is a distraction too.

2

The challenge

Automation saves effort but takes decisions away. In a car, a surprise on the screen is a distraction too.

3

The question

How do drivers experience suggestions they approve compared with actions the car applies by itself, in terms of predictability, control, trust and mental effort?

3

The question

How do drivers experience suggestions they approve compared with actions the car applies by itself, in terms of predictability, control, trust and mental effort?

The design

I designed Suggested Actions, a compact panel that appears when the driving context makes an action likely. To isolate the effect of automation, I built the same infotainment system three ways.

A. Manual

No suggestions. The driver finds every function in the menu, and every action is at most two taps away. This is the baseline.

Manual version: main menu, 4x2 grid

B. Advisory

The panel proposes an action and says why. The driver taps Apply, dismisses it, or ignores it and uses the menu.

Advisory version: Suggested Actions panel with Apply button over the map

C. Auto-applied

The system applies the action first and then confirms it in green. The driver only closes the panel.

Auto-applied version: green “Suggestions applied” panel

Design decisions

1

A stable layout underneath

The 4x2 menu never changes. Adaptive behavior lives only in the panel, so any surprise comes from the feature and not from a shifting interface.

1

A stable layout underneath

The 4x2 menu never changes. Adaptive behavior lives only in the panel, so any surprise comes from the feature and not from a shifting interface.

2

The panel sits on the driver’s side

The screen is mounted to the right of the wheel, so the panel is on the left edge, the shortest reach from the driver’s hand.

2

The panel sits on the driver’s side

The screen is mounted to the right of the wheel, so the panel is on the left edge, the shortest reach from the driver’s hand.

3

A short reason with every suggestion

Each panel says why it appeared (“School route detected”), using the familiar notification pattern of title, reason and one clear action.

3

A short reason with every suggestion

Each panel says why it appeared (“School route detected”), using the familiar notification pattern of title, reason and one clear action.

4

State you can read at a glance

A white Apply button means a suggestion is waiting. Green “Suggestions applied” means the car already acted. A dark theme and icon-first navigation keep glare and reading low.

4

State you can read at a glance

A white Apply button means a suggestion is waiting. Green “Suggestions applied” means the car already acted. A dark theme and icon-first navigation keep glare and reading low.

5

Never fully hands-off

In every scenario one task had no suggestion, so drivers kept using the menu, as they would in a real car that cannot anticipate everything.

5

Never fully hands-off

In every scenario one task had no suggestion, so drivers kept using the menu, as they would in a real car that cannot anticipate everything.

Prototype

I built 60 screens and 12 task flows in Figma and ran them on an iPad mounted where a center display would be. Two full pilot sessions showed that it was not always clear when a task was finished, so I added clearer completion feedback and return-to-menu transitions before the study.

School Drive

Navigate to school, turn on child lock, raise brightness, play kids music

Commute Home

Navigate home, switch drive mode to Comfort, lower brightness, play a podcast

Weekend Family Trip

Navigate to a picnic field, set steering feel to Light, set brakes to Light, play road trip music

All prototype screens for scenario 1
All prototype screens for scenario 2
All prototype screens for scenario 3

All 60 screens, organized by scenario and task.

How I did it

How I did it

1. Simulator setup

1. Simulator setup

A fixed-base driving simulator with a steering wheel and pedals, a town road at 50 to 60 km/h with oncoming traffic, and the prototype on an iPad to the right of the wheel. Participants kept driving while they used the screen.

A fixed-base driving simulator with a steering wheel and pedals, a town road at 50 to 60 km/h with oncoming traffic, and the prototype on an iPad to the right of the wheel. Participants kept driving while they used the screen.

Photo of the driving simulator with the iPad

2. Study design

2. Study design

12 licensed drivers each used all three versions. The order was fully counterbalanced: six possible orders, two drivers per order, so learning and fatigue were spread evenly.

12 licensed drivers each used all three versions. The order was fully counterbalanced: six possible orders, two drivers per order, so learning and fatigue were spread evenly.

3. Measures

3. Measures

After each version, a short questionnaire: task ease (SEQ), satisfaction (ASQ), and four items I wrote for predictability, control, trust and mental demand. After the last drive, a 10-minute interview.

After each version, a short questionnaire: task ease (SEQ), satisfaction (ASQ), and four items I wrote for predictability, control, trust and mental demand. After the last drive, a 10-minute interview.

4. Analysis

4. Analysis

Friedman tests with post hoc comparisons in JASP for the ratings. Reflexive thematic analysis in MAXQDA for the interviews.

Friedman tests with post hoc comparisons in JASP for the ratings. Reflexive thematic analysis in MAXQDA for the interviews.

Interview coding in MAXQDA

What I found

What I found

With 12 participants, none of the differences reached statistical significance, so I treat the numbers as direction, not proof. The direction was consistent, and the interviews explained it.

With 12 participants, none of the differences reached statistical significance, so I treat the numbers as direction, not proof. The direction was consistent, and the interviews explained it.

1. Fewer taps did not mean less effort

1. Fewer taps did not mean less effort

I expected mental demand to fall as automation went up. It did not. The auto-applied version needed the fewest taps, yet drivers rated it exactly as demanding as doing everything by hand (5.08 vs 5.08 out of 7). Advisory was the lowest (4.42) and also felt easiest (5.08 vs 4.00 for auto-applied). Drivers explained why: when the car had already acted, they had to work out what it did and whether it was right.

I expected mental demand to fall as automation went up. It did not. The auto-applied version needed the fewest taps, yet drivers rated it exactly as demanding as doing everything by hand (5.08 vs 5.08 out of 7). Advisory was the lowest (4.42) and also felt easiest (5.08 vs 4.00 for auto-applied). Drivers explained why: when the car had already acted, they had to work out what it did and whether it was right.

Manual

Advisory

Auto-applied

Mental demand

Mean rating, 0 to 7 axis (lower is better)

5.08

4.42

5.08

Manual

Advisory

Auto-applied

Task ease

Mean rating, 0 to 7 axis (higher is better)

4.58

5.08

4.00

Manual

Advisory

Auto-applied

“My mind was telling me ok everything has already been applied. So I think that felt a bit weird for me.”

Participant 1

2. Drivers wanted to keep the decision

2. Drivers wanted to keep the decision

Predictability, control and trust all fell step by step from manual to advisory to auto-applied. Control was the dominant theme in the interviews. Drivers were not asking to do more work. They wanted to stay the one who decides.

Predictability, control and trust all fell step by step from manual to advisory to auto-applied. Control was the dominant theme in the interviews. Drivers were not asking to do more work. They wanted to stay the one who decides.

“I have some control, but not like I’m doing everything manually.”

Participant 12, on the advisory version

3. An explanation after the fact is only a notification

3. An explanation after the fact is only a notification

Both adaptive versions showed the same one-line reason. In the advisory version it arrived before the decision and helped drivers understand the system. In the auto-applied version it arrived after the action, when it could no longer inform anything. Several drivers glanced, accepted and looked back at the road without checking. I call this compliant acceptance, as opposed to calibrated trust. It matters because a driver who does not check cannot catch a wrong action.

Both adaptive versions showed the same one-line reason. In the advisory version it arrived before the decision and helped drivers understand the system. In the auto-applied version it arrived after the action, when it could no longer inform anything. Several drivers glanced, accepted and looked back at the road without checking. I call this compliant acceptance, as opposed to calibrated trust. It matters because a driver who does not check cannot catch a wrong action.

“If it gives me consent I feel a bit more trust in the system.”

Participant 1

4. Some drivers never saw the feature

Engagement ranged from active use to not noticing the panel at all. Drivers who missed it simply used the menu and experienced every version as manual. Asked which version they would want in a real car, four drivers named advisory, four named manual, one said it depends on the task, and nobody named auto-applied.

“I think I didn’t even notice them, to be honest.”

Participant 9

Measure

Measure

Manual

Manual

Advisory

Advisory

Auto-applied

Auto-applied

Predictability

Predictability

5.92

5.92

5.33

5.33

4.92

4.92

Control

Control

5.25

5.25

4.67

4.67

4.33

4.33

Trust

Trust

6.25

6.25

5.83

5.83

5.42

5.42

Mental demand (lower is better)

Mental demand (lower is better)

5.08

5.08

4.42

4.42

5.08

5.08

Task ease (SEQ)

Task ease (SEQ)

4.58

4.58

5.08

5.08

4.00

4.00

Satisfaction with ease (ASQ)

Satisfaction with ease (ASQ)

5.08

5.08

5.00

5.00

4.42

4.42

Satisfaction with time (ASQ)

Satisfaction with time (ASQ)

4.83

4.83

4.75

4.75

4.17

4.17

Satisfaction with support (ASQ)

Satisfaction with support (ASQ)

5.75

5.75

5.17

5.17

4.25

4.25

Mean ratings, 1 to 7 scale, N = 12. No difference was statistically significant (Friedman tests). The strongest trends were predictability and control (both p = .082).

Design recommendations

Design recommendations

1. Suggest, don’t act

1. Suggest, don’t act

For actions that are not safety-critical, offer the action and let the driver confirm it with one tap.

For actions that are not safety-critical, offer the action and let the driver confirm it with one tap.

2. Explain before, not after

2. Explain before, not after

Put the reason in front of the decision. That is where it builds understanding and trust.

Put the reason in front of the decision. That is where it builds understanding and trust.

3. Make it discoverable

3. Make it discoverable

Introduce the feature on first use and add a subtle cue. A suggestion nobody notices helps nobody.

Introduce the feature on first use and add a subtle cue. A suggestion nobody notices helps nobody.

4. Add a voice or audio channel

4. Add a voice or audio channel

The most requested improvement in the interviews: hear the suggestion and keep your eyes on the road.

The most requested improvement in the interviews: hear the suggestion and keep your eyes on the road.

5. Match initiative to the stakes

Drivers welcomed suggestions for routine tasks like navigation and music more than for consequential ones like the child lock.

Wrapping it up

Wrapping it up

What I learned

The most useful result was the one that contradicted my hypothesis, and it only became explainable because I ran ratings and interviews together. For adaptive interfaces, the design question is less “how much can we automate” and more “where does the decision stay”.

What I would do next

Test with a larger and more varied group of 30 or more drivers to see whether the trends hold. Compare the visual panel with a version that adds an audio cue. Vary when and how the explanation appears. Use eye tracking to check whether drivers actually read the reason before accepting. I would also redesign the panel around icons, since several drivers said reading text while driving felt like extra load.

Limits: 12 participants from a university population, a scripted prototype that was not connected to the simulator, and self-reported experience only, with no glance or lane data.

Limits: 12 participants from a university population, a scripted prototype that was not connected to the simulator, and self-reported experience only, with no glance or lane data.

©2024 Kutay Yilmaz

©2024 Kutay Yilmaz