Effect of Sharedattention for HumanRobot Interaction

1
Effect of Shared-attention for Human-Robot
Interaction

• Junji Yamato
• jy_at_acm.org
• NTT Communication Science Labs., NTT Corp. Japan
• Kazuhiko Shinozawa, Futoshi Naya
• ATR Intelligent Robot and Communication Labs.

2
Aim

• To build Social Robot/Agent
• Sub goal
• To establish
• Evaluation methods
• Design guidelines
• for communication of human-robot/agent

3
Method

• To measure the influence of Agent/Robot on users
• Acceptance ratio of agent/robot recommendation

4
Color name selection task
Blue or Green? Cobalt green or emerald
green? Skin color or KARE-IRO? SUMIRE-IRO or
AYAME-IRO? ---- ---- Total30 questions. (from
color name text book)
5
Four experiments

• Compared agent and robot
• Compared agent and robot in physical world
• Measured the effect of eye contact
• Measured the effect of shared-attention

Detailed description of Experiment 1 and 2
Shinozawa, K., Naya, F., Yamato, J., and Kogure,
K. Differences in Effect of Robot and Screen
Agent Recommendations on Human Decision-Making ,
IJHCS (to appear)
Experiment 1, 2, and description of K4(robot)
Yamato, J., Shinozawa, K., Brooks, R., and Naya,
F. Human-Robot Dynamic Social Interaction. NTT
Technical Review 1, 6(2003), 37-43.
Available on-line http//www.ntt.co.jp/tr/ Back
number -gt Sep. 2003
6
Experiment 1Compare Agent and Robot
Agent
Robot
Agent
Robot
Conditions 30 questions, 30 subjects in each
group - Same question sequences, same voice,
similar gesture Measurement acceptance ratio,
questionnaire
7
Experiment 1 Robot
8
Experiment 1 Result
Acceptance agent gt robot (plt.01)
Familiarityindependent
9
Initial expectation
Robot has more influence because it lives in 3D
world, same as subjects.
agent
?
robot

10
Experiment 2 Compare in physical world
Color plate
Button box
Button box

• No recommendation (30 subjects)
• Recommendation by robot(31 subjects)
• Recommendation by agent (30 subjects)

11
Experiments
12
Experiment 2 Result
selection ratiorobot gt agent ( p lt 0.05)
robotgtgt no recommendation ( p lt 0.01)
13
Experiment 1 and 2 Results
Consistency matters.
Physical world
Media world
agent
robot
14
Why robot is better?

• Easy to detect gaze
• Eye contact
• Shared attention/joint attention

Measure the effect of eye contact and
shared-attention
15
Experiment 3 Effect of eye contact (mutual gaze)

• Eye contact was established by face tracking
• Eye contact time period that subject looked at
  robot and robot looked at subject
• Eye contact time and selection ratio?
• Two groups (14 subjects each)
• Eye contact, and NO eye contact

16
Robots
17
Selection ratio

• Higher selection ratio for eye contact group
• K4 No E.C. lt E.C. (p0.012)
• Rabbit No E.C. lt E.C. (p0.003)

18
Experiment 4 Effect of shared-attention

• Shared attention
• Period that robot looks at an object and subject
  looks at the same object. (color plate, button
  box)
• SA time and selection ratio
• Is there correlation?

19
Establishing shared-attention

• Robot looks at color plate and button box by
  prepared program
• Eye contact established by face tracking

Example video
20
Experimental conditions

• 28 subjects
• SA time 51.7 sec (total for 30 questions)
• (Longer than in Experiment 3 )
• Selection ratio. Average 0.57 S.D. 0.14
• Some subjects were positive, and others were not.
  Clear contrast, from the questionnaire.
• Example Robot is prompting wrong choice. I
  feel the robot forced me to select his
  recommendation (negative).

21
SA time and selection ratio

• No correlation

Selection ratio
Shared-Attention time (count) 50count1sec.
22
Clustering subjects by TEG(Ego-gram)

• Ego-gram based on transactional analysis
• Measure three ego-states by questionnaire
• CP, NP (critical parent, nurturing parent)
• A (adult)
• FC, AC (free child, adapted child)
• TEG (Tokyo Univ. Egogram)is common in Japan

23
High/Low TEG measurement and SA time.

• Strong correlation in SA time and acceptance
  ratio for high AC (Adapted Child) group

24
SA time and selection ratio (high AC low CP
group)

• Positive correlation(Speamans r0.51,p0.051).

25
SA time and selection ratio

• High-SA group high selection ratio (plt0.05)

(high AC group)
26
Result and Discussion

• High AC subject (obedient type) showed positive
  correlation between SA time and selection ratio.
• No significant difference between SA time itself
  and selection ratios for high AC and low AC
  groups
• Eye contact and shared-attention promote close
  communication. Some people like such intimate
  relation, but others dont. It depends on the
  character.
• SA is effective. Even SA was not actually
  realized. We do not need to develop image
  understanding technology we just have to fake it.