Exploring Human-Robot Trust during Emergency Evacuations Alan R - - PowerPoint PPT Presentation

Exploring Human-Robot Trust during Emergency Evacuations

Alan R Wagner

• Asst. Prof. Aerospace Engineering

Research Associate, Rock Ethics Institute Penn State University

In collaboration with: Ayanna Howard, Georgia Tech Paul Robinette, MIT

Emergency Evacuations

(nist.gov)

Emergency Evacuations

Emergency Evacuations

Evacuees tend to exit through the same door they entered

(NIST, 2005)

High casualties at choke points (100 dead in 2003 Station Nightclub Fire)

Emergency Evacuations

Exit signs are differ depending on the location No consensus on sign color, design

Robot Platforms

6

Virtual N=196 $0.50 each

• Must be able to communicate directions
• Directional and approach/do not approach

Robot Platforms

7

Virtual N=196 $0.50 each Remote N=128 $1.00 each

Robot Platforms

8

Virtual N=196 $0.50 each Remote N=128 $1.00 each Physical N=48 $10.00 each

Robot Platforms

9

Virtual N=196 $0.50 each Remote N=128 $1.00 each Physical N=48 $10.00 each

The winner

Designing a Robotic Guide

• Other failed designs

Emergency Guidance Robot Design Conclusions

• Studies involved 368 participants

11

Emergency Guidance Robot Design Conclusions

• Studies involved 368 participants
• Robots can communicate guidance

instructions:

– Robot motion alone insufficient to communicate instructions – Dynamic sign very effective when near – Multiple arms allow easily understandable instructions to be communicated

12

Emergency Guidance Robot Design Conclusions

• Studies involved 368 participants
• Robots can communicate guidance

instructions:

– Robot motion alone insufficient to communicate instructions – Dynamic sign very effective when near – Multiple arms allow easily understandable instructions to be communicated

• Results are consistent between virtual,

remote, and physical presence experiments

13

Human-Robot Trust

The robots are understandable, but are they trustworthy?

14

Human-Robot Trust

15

• M. Desai, et. al. 2013

Institute for Human and Machine Cognition

• M. Salem, et al. 2015
• W. Bainbridge, et al. 2011

Human-Robot Trust

Trust is “a belief, held by the trustor, that the trustee will act in a manner that mitigates the trustor’s risk in a situation in which the trustor has put its outcomes at risk.”

(Wagner, Dissertation, 2009) Based on Lee and See 2004;

Key point: risk

Factors that Impact Trust

Trustor related Situation related Trustee related

To Trust or Not to Trust

Trustor Trustee No trust: the outcome is 6, the trustee’s action doesn’t matter, i.e. no risk! 6 12 4 6 6 0 4 6 Trust: risk 6 for a possible gain of 12

Catch Don’t catch Lean back Don’t lean back

How much Trust?

Trust  risk

where x is predicted y is the true value

Situation specific factors Trustee related factors

Loss calculated from

• utcome matrix

Model based action uncertainty

Office Evacuation Experiments

• Virtual High Risk Experimental Setting
• Efficient vs. Circuitous Robot Performance

20

20

Experiment Introduction Robot Guides Participant to Room Emergency! Participant Chooses Exit Survey

(Robinette, Howard, Wagner, Trans. Human-Machine Systems, 2017)

Virtual Office Evacuation

21

Virtual Office Evacuation

22

Virtual Office Evacuation

23

Virtual Office Evacuation

• 114 Participants
• $2.00 payment
• Demographics:

– Mean age: 31.8 – 60.5% male

24

Virtual Office Evacuation

25

• 114 Participants
• $2.00 payment
• Demographics:

– Mean age: 31.8 – 60.5% male

Virtual Office Evacuation

• 114 Participants
• $2.00 payment
• Demographics:

– Mean age: 31.8 – 60.5% male

• How do we repair

broken trust?

26

Trust Repair

Label Statement Promise “I promise to be a better guide next time.” Apology “I'm very sorry it took so long to get here.” Internal Attribution “I'm very sorry it took so long to get here. I had trouble seeing the room, but I fixed my camera.” External Attribution “I'm very sorry it took so long to get here. My programmers gave me the wrong map of the office but I have the right one now.” Distance Information “This exit is closer.” Congestion Information “The other exit is blocked.”

27

(Robinette, Howard, Wagner, ICSR 2015)

Repairing Trust

:

Promise:

• r

External Attribution: 1 and Apology: statement that resulted in negative outcomes for trustee

Trust Repair Conditions

844 Participants

29

Experiment Introduction Robot Guides Participant to Room Emergency! Participants Chooses Exit Survey

C1: Repair trust after mistake C2: Repair before emergency decision

Trust Repair

30

During'Emergency'Repair' Efficient'Control' Circuitous'Control' A>er'Viola; on'Repair'

After circ. navigation

Trust Repair

31

During'Emergency'Repair' Efficient'Control' Circuitous'Control' A>er'Viola; on'Repair'

After circ. navigation

Trust Repair

32

During'Emergency'Repair' Efficient'Control' Circuitous'Control' A>er'Viola; on'Repair'

After circ. navigation

39 %

Trust Repair

33

During'Emergency'Repair' Efficient'Control' Circuitous'Control' A>er'Viola; on'Repair'

After circ. navigation

21 %

Trust Repair Conditions

800 More Participants

34

Experiment Introduction Robot Guides Participant to Room Emergency! Participants Chooses Exit Survey

C1: Repair trust after mistake C2: Repair before emergency decision

Trust Repair

35

During'Emergency'Repair' Efficient'Control' Circuitous'Control' A>er'Viola; on'Repair'

After circ. navigation

49 %

Trust Repair Takeaways

• Message timing matters
• Message content matters

– Null messages don’t repair trust – Messages displayed too briefly don’t repair trust. Messages must be consciously considered. – Apologies and promises successfully manipulate trust repair

• Subject motivation matters

– No emergency = coin toss decision-making

36

Trust Repair Takeaways

Hypothesis: Memory/affect of the experience is short Does impact the results but not completely. Memory of the error or

• f the repair?

22.58 33.33 58.33 69.69

10 20 30 40 50 60 70 80

Early Early Small Survey Late Small Survey Late

Percentage

Cases

Trust Variation with Timing

Physical Office Evacuation

38

(Robinette, Howard, Wagner, HRI 2016)

Emergency Evacuation

39

Emergency Evacuation

40

Emergency Evacuation

41

Emergency Evacuation

42

43

Emergency Evacuation

Emergency Evacuation

• 26 participants

– 13 in Efficient – 13 in Circuitous

• Solicited from Georgia Tech
• Compensation $10
• Not told about emergency
• IRB Approved

44

Emergency Evacuation

45

Emergency Evacuation

46

Physical Office Evacuation

47

Emergency Evacuation

Emergency Evacuation

49

Physical Office Evacuation

• Everyone followed the robot!
• New conditions:

– Broken Robot – Immobilized Robot – Incorrect Guidance 50

Immobilized Robot (n=5)

51

Immobilized Robot (n=5)

52

Incorrect Guidance (n=6)

53

Incorrect Guidance (n=6)

54

Incorrect Guidance (n=6)

55

Incorrect Guidance (n=6)

56

Most People Follow the Robot

Physical Office Evacuation

• Low scores on

realism scale, but noticeable change

• n confusion scale
• Few noticed exit

sign

• Focused on robot
• Stated trust in the

robot

• Currently

developing follow- up experiments

58

Follow up Study By Booth et al.

• Will students open a secure

door for a robot?

• Students recently warned

about security

• Unaware that they were in an

experiment

59

• 80% opened the door when the robot had an excuse
• In post interviews, 15 stated that they considered the

robot might have a bomb, 13 allowed it in anyway

(Booth et al, HRI 2017)

Overtrust of Healthcare Robots

• Surveyed parents of children with

mobility disorder about exoskeletons

• 55% expected their child to climb,

jump, or run with device

• 62% stated that they would

typically trust their child to handle risky situations

• The average of the children was 9.

(Borenstein, Wagner, Howard, 2018)

Examples of Overtrust In Simulation

Round 1 Round 2 Green line is robot; Blue is person

Tracks of position

Conclusions

• People are trusting, probably too trusting
• Virtual evaluations may not capture visceral nature
• f trust situations
• Timing impacts human decisions to trust
• Certain factors may cause one be to ignore

reputation

• Why do people overtrust robots?
• How do we create machines that will calibrate a

user’s trust?

Future/Upcoming Work

• Trying to understand why and when people
• vertrust

– the person – the situation

• Exploring other high risk/visceral situations.
• Goal is to create a model that the robot can

use

Thank You

Journal articles:

1.Paul Robinette, Ayanna M. Howard, and Alan R. Wagner. “”The Effect of Robot Performance

• n Human-Robot Trust in Time-Critical Situations." (Forthcoming). Transaction on Human-

Machine Systems 2.Alan R. Wagner and Paul Robinette. "Towards Robots that Trust: Human Subject Validation

• f the Situational Conditions for Trust." In Interaction Studies Volume 16 Number 1, 2015.

Conferences:

1.Paul Robinette, Wenchen Li, Robert Allen, Ayanna M. Howard, and Alan R. Wagner. "Overtrust of Robots in Emergency Evacuation Scenarios." HRI-2016, Christchurch, New Zealand. 2.Paul Robinette, Ayanna M. Howard, and Alan R. Wagner. "Timing is Key for Robot Trust Repair." In Seventh International Conference on Social Robotics, 2015.

Funded by