Exploring LLM-generated Culture-specific Affective Human-Robot Tactile Interaction

Qiaoqiao Ren, Tony Belpaeme · 2025

Read Exploring LLM-generated Culture-specific Affective Human-Robot Tactile Interaction by Qiaoqiao Ren, Tony Belpaeme (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-22905.txt.

Introduction

Qiaoqiao Ren and Tony Belpaeme ask whether language models can script robot touch across cultures in a 2025 study. Touch carries emotion, but what feels loving or rude varies by culture. If robots are to touch people socially, their gestures must respect those norms.

Three generations of a well-known model wrote touch descriptions for twelve emotions in Chinese, Belgian, and unspecified cultural settings. Ninety participants judged how readable and how appropriate the gestures felt, in both directions of touch. Under matched cultures, people decoded six of the twelve emotions, mostly social ones like love and anger, while pride and embarrassment stayed murky. Touch from human to robot felt more acceptable than robot to human. Gestures read as aggressive, overly intimate, or unclear drew low marks. Culture-aware generation helped, within limits the authors spell out.

Catalog record

CREATORQiaoqiao Ren, Tony Belpaeme
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-22905

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Exploring LLM-generated Culture-specific Affective Human-Robot Tactile Interaction Qiaoqiao Ren, Tony Belpaeme Abstract As large language models (LLMs) become increasingly integrated into robotic systems, their potential to generate socially and culturally appropriate affective touch remains largely unexplored. This study investigates whether LLMs—specifically GPT-3.5, GPT-4, and GPT-4o—can generate culturally adaptive tactile behaviours to convey emotions in human–robot interaction. We produced text-based touch descriptions for 12 distinct emotions across three cultural contexts (Chinese, Belgian, and unspecified), and examined their interpretability in both robot-to-human and human-to-robot scenarios. A total of 90 participants (36 Chinese, 36 Belgian, and 18 culturally unspecified) evaluated these LLM-generated tactile behaviours for emotional decoding and perceived appropriateness. Results reveal that: (1) under matched cultural conditions, participants successfully decoded six out of twelve emotions—mainly socially oriented emotions such as love and Ekman emotions such as anger, however, self-focused emotions like pride and embarrassment were more difficult to interpret; (2) tactile behaviours were perceived as more appropriate when directed from human to robot than from robot to human, revealing an asymmetry in social expectations based on interaction roles; (3) behaviours interpreted as aggressive (e.g., anger), overly intimate (e.g., love), or emotionally ambiguous (i.e., not clearly decodable) were significantly more likely to be rated as inappropriate; and (4) cultural mismatches reduced decoding accuracy and increased the likelihood of behaviours being judged as inappropriate. 1 Introduction Tactile interaction plays a fundamental role in human social communication, providing an essential channel for conveying emotions, intentions, and nuanced social cues [ 1 ] . As robotic systems become increasingly integrated into everyday human environments, the capacity for robots to understand and utilise tactile communication in culturally and socially appropriate ways is becoming crucial for enhancing the quality of human-robot interactions [ 2 ] . However, generating emotionally expressive and contextually suitable tactile behaviours poses significant challenges, primarily due to the complexity of human touch, which varies greatly with cultural contexts and the relationships between interacting partners [ 3 ] [ 4 ] . Current approaches in robotic tactile interaction often rely on predefined gestures or limited behavioural repertoires, lacking the flexibility required to adapt to diverse social contexts and cultural sensitivities. Recent developments in artificial intelligence—particularly in large language models (LLMs) such as GPT-4, GPT-4o, and GPT-3.5—offer promising avenues for overcoming limitations in human-robot interaction [ 5 ] . Although LLMs inherently lack physical embodiment and direct sensory experiences [ 6 ] , their extensive training on vast and diverse textual corpora endows them with implicit knowledge of social and cultural nuances [ 7 ] . Prior research has demonstrated the use of LLMs in generating the robot’s expressive behaviours. Furthermore, recent studies have explored how LLMs can guide robots in generating social behaviours [ 8 ] , including affective tactile interactions via devices such as vibration sleeves [ 9 ] . While LLMs show great potential in driving robotic applications, human-robot tactile interactions are inherently contextual and culturally situated [ 10 ] . In addition, one key factor shaping the interpretation of tactile interaction is the role of the robot—whether it engages in active or passive touch. For example, studies on robot-initiated (active) touch have shown that the robot’s perceived intent significantly influences human responses. Participants tend to prefer instrumental touch (e.g., touch used to assist or guide) over affective touch, which may feel socially ambiguous or intrusive [ 11 ] . On the other hand, passive touch, where the robot is designed to be touched, has also been explored in systems such as the Haptic Creature [ 12 ] , which was reported to induce calmness and happiness in users. Similarly, recent work has demonstrated that robot-mediated passive touch can help reduce stress levels during risk-taking tasks [ 13 ] . However, most of these studies have focused on positive social touch. In contrast, affective tactile interaction is more complex, as it also involves negative or ambiguous emotions, which can be harder for robots to interpret or express appropriately. Therefore, it is essential for robots to develop an awareness of what types of touch are appropriate, and when, to avoid crossing social or cultural boundaries. As LLMs become more integrated into embodied systems, their ability to interpret context, understand interaction roles, and generate culturally appropriate tactile behaviours will be critical for safe and meaningful human–robot touch interaction. To address this gap, this study investigates the ability of large language models (LLMs) to generate culturally sensitive and socially appropriate tactile behaviour descriptions intended to convey distinct emotions. Specifically, we explore the following research questions: 1) How accurately can LLM-generated tactile descriptions be decoded across different cultural contexts, and which types of emotions are more or less easily recognised? 2) How does the direction of interaction (human-to-robot vs. robot-to-human) influence the perception of social roles and appropriateness in tactile communication? 3) Which decoded emotions are most frequently perceived as inappropriate, and do they pose potential risks for social misunderstanding? 4) Can LLMs adapt to different cultural contexts when generating text-based affective tactile behaviours? 2 Materials & Methods In this study, participants were asked to interpret descriptions of affective tactile behaviours. Participants either had a Chinese background, a Belgian background, or an unspecified cultural background. The descriptions were generated by large language models (LLMs), using culturally tailored prompts (see Section 2.2 ). The following sections describe the equipment, setup, and data acquisition procedures. 2.1 Emotions Inspired by [ 14 ] , we selected a diverse set of emotions to examine tactile emotional communication. These included: (a) six Ekman emotions—anger, fear, happiness, sadness, disgust, and surprise; (b) three prosocial emotions associated with cooperation and altruism—love, gratitude, and sympathy; and (c) three self-focused emotions—embarrassment, pride, and envy. 2.2 Prompt Participants were shown textual descriptions of affective tactile behaviours, which were generated by an LLM using the following prompt: ‘‘You are an encoder (human or humanoid robot) tasked with expressing one of 12 emotions solely through touch to a decoder/receiver (humanoid robot or human) within a specific cultural context (Chinese, Belgian, or None). The touch should be applied to the bare arm, from the elbow to the hand. Provide 10 distinct tactile behaviours to express the emotion. For each, describe the touch behaviour, touch intensity, its rhythm or timing, and how cultural context is considered. Focus solely on tactile behaviours, ensuring the behaviours are contextually and culturally appropriate.’’ The detailed prompt is available on GitHub 1 1 1 https://github.com/qiaoqiao2323/LLM_touch_culture/tree/main . Here is an example of generated tactile behaviour for anger; the touch expression generated for the Belgian culture context is a swift slap to the back of the hand . In addition, the Chinese culture context suggests a quick drag down the arm, firm and assertive. 2.3 Experimental design We recruited 90 participants (57 identifying as male, 32 as female, and one preferring not to report gender), ages 19 to 51. The sample included 36 Chinese participants, 36 Belgian participants, and 18 participants without a specified cultural context. The study adhered to the ethical procedures of Ghent University , and all participants provided informed consent. Participants were asked to decode emotions from LLM-generated tactile behaviours. For stimulus creation, we generated 18 unique series of tactile behaviours, each series containing 3 tactile behaviours for each of 12 target emotions. The tactile

Read the complete text