ReTracing

Abstract

ReTracing is a multi-agent embodied performance art adopting an archaeological approach to examine how artificial intelligence shapes and produces bodily movement. Drawing from science-fiction novels, large language models generate paired prompts—“what to do” and “what not to do”—which diffusion-based models transform into choreographic guides for human performers and motor commands for quadruped robots. The mirrored stage and motion tracking create a 3D digital archive of movement, reflecting how generative systems encode socio-cultural biases through choreographed embodiment.

Framework

ReTracing integrates text, motion, and embodiment through three key components:

• Excavating Encoded Bodies: Using LLMs (e.g., Qwen-2.5) to translate literary excerpts into movement prompts.
• Embodied Translation: Text-to-video diffusion models transform prompts into visual choreography for human and robotic agents.
• Digital Archaeology: Multi-camera 3D reconstruction records and analyzes motion traces, revealing how AI internalizes embodiment.