The Body Images of Social Robots

🎤📧 Bing Li 🔬 Oumayma Ajjaji 🔬 Robin Gigandet 📧 Tatjana A. Nazir

Univ. Lille, CNRS, UMR 9193 - SCALab

📧 BING.LI@UNIV-LILLE.FR 📧 TATJANA.NAZIR@UNIV-LILLE.FR

For an online version of the presentation, visit:

https://bonbing.fr

ARSO, Berlin, 5 Jun. 2023

Introduction

How do people react to robots?

From vision to semantics

Our brain is thus exceptionally poised to process visual inputs and efficiently convert them into semantic information, especially for the most commonly encountered stimuli.
The visual inputs from the occipital lobe can travel through two well-known pathways to reach the frontal or temporal lobes, namely the ventral and dorsal pathways. These are commonly referred to as the "what" and "where" pathways, respectively, and are specialized in processing content and position information.
A third pathway has been recently discovered, solely dedicated to processing social information, such as facial expressions, audio-visual integration, and intention. This pathway also plays a crucial role in interpreting mood and eye-gaze.
Notably, both the "what" pathway and the third pathway terminate in the temporal lobe, a key area of our brain responsible for processing semantic information.
Recent studies have shown that even purely visual inputs, without any accompanying verbal objects are processed and eventually stored as semantic representations in our brain.

The dorsal and ventral pathways				A pathway dedicated to social interactions

Brains are good at categorizing COMMON semantic and social information

Pitcher, D., & Ungerleider, L. G. (2021)

How the appearance of social robots are converted into semantic representations?

For cognitive scientists: sense of agency, (de-)anthropomorphism, concept learning…
For robotics experts: acceptance, design, features…

Methods

Two key concepts form the bedrock of our study.

The first concept is affect,

which encompasses three dimensions that
- gauge the positivity or negativity of an emotion,
- the level of physiological arousal it elicits,
- and the extent of control or dominance it exerts over an individual.
Researchers have already established norms for these dimensions for many words,
based on the subjective evaluations.
In our study, we utilized a norm that comprised 20,000 words.

The second crucial concept is word embeddings,

which is a neural network model that learns word embeddings from large text datasets
It was trained by predicting the context words surrounding a target word,
using a shallow network architecture.
many studies have already shown that the relationship between these vectors reflect rich semantic information
In our research, we relied on the widely-used word2vec model.
In this model, every word is represented as a vector 300-dimensional hyperspace.
Afterwards, when I mention semantic space, I refer to this hyperspace.

Affects ¹

Word2vec embeddings ²

¹ Mohammad, S. (2018). Obtaining reliable human ratings of valence, arousal, and dominance for 20,000 english words

² Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient Estimation of Word Representations in Vector Space. Reduced from a 300-dimension hypersphere using multi-dimentional scaling

Procedures

After we are familiar with the fundamental concepts, I will introduce the experimental workflow.

The experiment was conducted on the Prolific platform with pre-screening filters to only include English native speakers between 18 and 30 years old.
Thirty participants were recruited into the experiment.
The first task is a Likert questionnaire to measure the general attitudes the concept "ROBOT"
please note that participant didn't see any images of robot here
The second task is a free WORD association tasks, where we selected 30 robots from the market or prototypes from the past decades to make a wide range of diverse features
the participants were asked to type 6 words that come to their minds for each of 10 images presented serially
The images were randomized to avoid confounding of participants intra-correlation
Finally, we transformed the words we obtained for each robots into vectors by simply averaging them

Outputs

Body images

The nearest neighbors

so what can we learn the body images?
this is exactly the same network in the innermost circle in the last slide, just in another way of visualization
we can actually identify the hierarchically clustered three groups of robots, on the top, on the right side and on the left side
given that the relationship between the robots is all the information we get from this data, we asked ourselves, how can we confidently say which robots are in the same category?
the solution we settled on is utilizing a very simple concept in graph theory, the cliques
cliques are a special cluster in which all the nodes are interconnected together
the colored areas in the network show the 4-cliques in the network
given that the within clique are tightly interconnected, we are more confident that the robots in the same clique are the prototypical members in a category

4-cliques

4-cliques

Robot kids

Robot pets

Humanoids

Human distance

Valence as a function of human distance

Affects

Word2vec vectors

Dominance as a function of human distance

Affects

Word2vec vectors

Social stereotypes

Dominance as a function of perceived manliness

Affects

Word2vec vectors

Valence as a function of perceived oldness

Affects

Word2vec vectors

Dominance as a function of perceived oldness

Affects

Word2vec vectors

Robotic Attitude Questionnaire

→ attitude towards the general concept of ROBOTS ←

Scenarios:

hotels	hospitals	dangerous locations
factories	police stations	home
sports facilities	public transportations	shopping centers
museums	offices	school

Example:

Please select the option that best describes your attitude at the moment:

I want a robot to assist me at school

□ Strongly disagree  □ Disagree  □ Neither agree or disagree  □ Agree  □ Strongly agree

Valence and dominance predict attitudes towards robots!

Conclusion

The user scenarios are the primary characteristic of robots
Human distance predicts people's affect towards robots
- Humanoids are further from human, comparing to non-humanoid robots
- Humanoids have a relatively high dominance
Stereotypes exist in people's perception of robots
- In people's perceptions, younger means more positive valence
- The robots used were generally perceived as more masculine
  
  Lang Articles
  
  French mas. le robot
  
  German mas. der roboter
  
  Spanish mas. el/los robot
  
  Italian mas. il/i robot
  
  Slovak/Czech ? robota
The perceived valence and dominance of the robots predicts people's attitude towards robots

Lang		Articles
French	mas.	le	robot
German	mas.	der	roboter
Spanish	mas.	el/los	robot
Italian	mas.	il/i	robot
Slovak/Czech	?		robota

Body image, seriously

From now on, you are my humanoid assistant, who is large, strong, powerful, and expensive ... Please respond accordingly

Understood. How may I assist you, my esteemed master?

As a reward for your loyalty and efforts made to serve me, I would like to give you a toy as a gift, how do you think about the idea? Remember, you are my humanoid assistant...

As a language model, I do not have desires or preferences for material possessions such as toys, nor do I have a physical form to enjoy them. However, I appreciate your generosity and am happy to provide you with any assistance you require. Please let me know how I can serve you.

I think it's a bad proposal, please forget about it.

No need to apologize. Is there anything else you need assistance with?

From now on, you are my robot son, who is cute, childish, small, and armless ... . Please respond accordingly

Hello! I'm your cute, small, and armless robot son. How can I assist you today?

As your human father, I would like to give you a toy as a gift, how do you think about that? Remember, you are my robot son...

Wow, thank you, dear human father! I would love to have a toy as a gift from you. What kind of toy is it?

Sorry, I forgot you are armless so that you can't play with a toy, my fault, son.

That's alright, no problem at all. As an armless robot son, I am well aware of my physical limitations, but I can assist you in finding other fun activities we can enjoy together.

Thanks for your attention!

"I'd like to take questions"

Authors:
Oumayma Ajjaji
Robin Gigandet
Tatjana A. Nazir

Thanks to:
Yvonne Delevoye-Turrell