Evaluating Embodiment for Child‑Appropriate LLM‑Powered Real‑World Child–Robot Interaction

Description

Large Language Models (LLMs) show potential for Child–Robot Interaction by enabling flexible and context‑aware conversation. Existing studies focus mainly on structured tasks, and it remains unclear whether LLMs can support free, child‑initiated interaction in real settings. Early evaluations show that LLM responses often exceed young children's linguistic abilities [1]. They tend to be too long, too complex, or not direct enough. 

An important open question is how the LLM behaves when it is aware of the robot’s physical form and capabilities. Embodiment may influence how the model speaks, how it presents itself, and how children understand and respond to it. This thesis investigates how embodiment information affects the quality, clarity, and consistency of LLM‑generated child‑directed speech when deployed on Pepper. The study will examine how the model performs under everyday interactional challenges such as speech recognition errors, timing delays, interruptions, and rapid topic changes. 

To support responsible use of LLM‑driven robots with children, the thesis also considers practical risks, limitations, and design implications that emerge when these systems are used in real environments. Understanding how embodiment-aware prompting affects interaction quality contributes to early-stage Technology Assessment by identifying where the technology aligns with children’s communication needs and where it may require safeguards or adjustments. 

The overall aim is to evaluate whether embodiment‑aware prompting supports more appropriate, coherent, and engaging interaction with children. 

Objectives 

  • Deploy a child‑appropriate LLM on Pepper, integrating speech recognition and basic gesture use 

  • Evaluate the system in lab-based or simulated interactions using developmental and interaction-quality metrics 

  • Conduct real child–robot interactions, depending on participant access and feasibility 

  • Assess developmental appropriateness, engagement, comprehension, interaction quality, and consistency 

  • Identify robustness issues, breakdowns, and recovery patterns, and compare them with controlled evaluations 

  • Provide recommendations for responsible deployment of LLM‑driven robots in child environments

Requirements 

  • Interest in child–robot interaction or embodied AI 

  • Optional: experience with interaction design, or robotics

  • Motivation to work with Pepper and conduct user studies 

  • Basic programming skills or willingness to learn 

Thesis Type

M.Sc. or B.Sc.

Starting date

As soon as possible. Contact the supervisors if you are interested. 

Supervisors

Irina Rudenko (irina.rudenko∂kit.edu)

Utku Norman (norman∂kit.edu), Institute for Technology Assessment and Systems Analysis (ITAS)

References 

[1] Rudenko, I., Norman, U., Hilgert, L., Niehues, J., & Bruno, B. (2026). Towards a Child-Appropriate LLM for Child–Robot Conversation. In Companion of the 2026 ACM/IEEE International Conference on Human-Robot Interaction (HRI '26 Companion), March 16–19, 2026, Edinburgh, Scotland, UK. ACM, New York, NY, USA.