Social Understanding for Android Robots

Impulse talk - IRIS general assembly

Prof. Dr. C. Becker-Asano, Hochschule der Medien Stuttgart

20.11.2025

What are android robots?

Example: Geminoid HI-6, Osaka EXPO, August 2025

Android Robots:

  • are a subclass of humanoid robots, which are a subclass of social robots, or:
    social > humanoid > android
  • are (mostly) designed indistinguishable from humans
  • feature very realistic facial expressions
  • are not mobile (most of them)
  • interact socially with humans

A laughing Geminoid HI-1 (2009)

The Ultimatum Game and a laughing android

Laugther in HRI (teleoperated)

  • is laughter in combination with robot gaze interpreted as “laughing at” the player?
    $\rightarrow$ not in this case
  • does robot laughter change the impression of a robot?
    $\rightarrow$ only slightly to the negative (i.e., more irritating)

Lesson learned: A scary robot remains a scary robot - even when laughing!

The risks of android robot design

The uncanny valley hypothesis and social robots (MacDorman & Ishiguro, 2006)

The Uncanny Valley (Mori, 1970)

  • More pronounced when artefact is moving
  • The more humanlike the design the more familiar we feel with it, until…
  • … the expectations are not fulfilled. Then the robot appears uncanny.

2023: Android Andrea in the Mercedes Benz Museum

  • ChatGPT 3.5
  • simple system prompt
  • only English language output
  • push-to-talk button

Three conditions:

  1. male voice, short hair
  2. female voice, short hair
  3. female voice, long hair

2023: Usefulness of Android Andrea

On a scale from 0 to 10 how usefull was Andrea today?

(between-subjects, N=42)

2025: Android Andrea again in the Mercedes Benz Museum

  • ChatGPT 4.1 with Retrieval Augmented Generation (RAG)
  • WASABI/ChatGPT emotion simulation
  • multiple languages with gender-neutral voice

Three conditions:

  1. no emotions (ChatGPTpure)
  2. ChatGPT-generated emotions (ChatGPTemotion)
  3. WASABI-generated emotions (ChatGPTemotion)

1. Preceived usefulness

Technology Acceptance Model (TAM): Perceived usefulness

(between-subjects, seven point scale, N=73)

Emotions make Android Andrea less usefull?

2. Preceived ease of use

Technology Acceptance Model (TAM): Perceived ease of use

(between-subjects, seven point scale, N=73)

Emotions make Android Andrea less easy to use?

Technical challenges with an android robot from Japan

Five robot heads for rapid development

  • basic emotions as facial expressions
  • lip-sync facial animation
  • people tracking through external camera
  • easier deployment than Android Andrea

Fourteen actuators for head movments

  • three for moving the head
  • two for the eyes
  • two for the upper and lower eyelids
  • two for the eye brows
  • three for the mouth
  • one for the chin

Lip-sync facial animation - mapping the face

PhD student Marcel Heisler

  • 2024: Mapping from Action Units to actuators (HRI conference, LBR)
  • training a multilayer perceptron through random motions using OpenFace 2.0

2025: “An iPhone is all you need”

Process explanation

Evaluation in three public places in Stuttgart

2025: Usefulness of robot head KIM

On a scale from 0 to 10 how usefull was robot head KIM today?

Social Understanding

Social Understaning through external sensors

Android ERICA - ATR, Kyoto, Japan, 2023

Backup

Emotional speech synthesis and animation

  • using emoknob “Enhanced Voice Cloning with Fine-Grained Emotion Control”
  • work in progress

Stationary android head with mobile assistant robot TEMI 3

We are planning to combine the robot heads with mobile robots for more flexible tour guiding through a museum.