敵か味方か？人間とAIにおける 「社会的意図」認識のギャップを解明

Distinguishing friendly from hostile intentions is a critical survival skill, particularly when verbal or facial cues are absent. While bodily emotion recognition is well-studied, the movement principles underlying “Friend vs. Foe” social intentions are under-explored. This paper presents a systematic study of embodied social intention from three perspectives: kinematic encoding, cross-cultural human expression and perception, and machine decoding. We constructed a novel motion capture database of 160 full-body recordings from 39 Taiwanese and 41 Japanese professional performers. An “imaginary alien” paradigm was used to elicit intention-driven movements while minimizing cultural stereotypes. Our analysis revealed a complex relationship between physical dynamics and social interpretation. Motion Energy Analysis (MEA) revealed that hostile intentions exhibit significantly higher energy concentrations in low-frequency bands than those in friendly intentional movements, with frequency-specific distinction between Japanese (0.25−1 Hz) and Taiwanese (0.5-1 Hz). Perceptual experiments with 77 observers from 3 cultural groups (Japanese, Taiwanese, Chinese) demonstrated robust cross-cultural consistency, with inter-group correlations exceeding rs>0.82. Qualitative analysis indicated that this human consensus relies heavily on inferring perceived mental states and movement dynamics. In contrast, a Spatio-Temporal Graph Convolutional Network (ST-GCN) achieved reliable classification accuracy but showed weak correspondence with human judgments (r=0.26). This discrepancy suggests that while current deep learning models can detect objective kinematic features of friendliness/hostility, they fail to capture the high-level cognitive inference strategies humans use to decode social intent. This work provides a challenging novel dataset and new insights into embodied social intention, and highlights key challenges for human-aligned social motion understanding.