💻 computer science

ReTracing: An Archaeological Approach Through Body, Machine, and Generative Systems

ReTracing is a multi-agent embodied performance art project that utilizes large language models and diffusion-based video generation to translate science-fiction narratives into choreographed movements for a human and a robot, creating a digital archive of motion traces to critically examine how generative AI encodes socio-cultural biases and shapes the definition of humanity in an age of intelligent machines.

Original authors: Yitong Wang, Yue Yao

Published 2026-02-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yitong Wang, Yue Yao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are an archaeologist, but instead of digging through dirt to find ancient pottery or bones, you are digging through code and stories to find out how computers "think" about human bodies.

That is exactly what the project ReTracing does. It's a fascinating art experiment that mixes science fiction, robots, and dance to ask a big question: When AI tells a human and a robot how to move, what does that say about who we are?

Here is a simple breakdown of how it works, using some everyday analogies:

1. The Excavation: Digging into Stories

The researchers started by picking scary, emotional, or strange stories from famous sci-fi novels (like Frankenstein or The Handmaid's Tale). These stories often describe people or creatures moving in ways that feel controlled, frightened, or mechanical.

The Analogy: Imagine these stories are like ancient scrolls. The researchers didn't just read them; they asked a super-smart AI (a Large Language Model) to act like a translator. They asked the AI: "If this character is feeling this way, what specific moves should they make? And what moves should they never make?"

2. The Choreography: The "Do" and "Don't" List

The AI took those story snippets and created two lists for every scene:

The "Do" List: A set of instructions on how to move (e.g., "Crouch down," "Tremble," "Walk slowly").
The "Don't" List: A set of instructions on what to avoid (e.g., "Don't smile," "Don't run fast").
The Analogy: Think of this like a strict dance instructor who is also a bit of a control freak. They aren't just telling the dancers what to do; they are also telling them what not to do, effectively trapping the movement in a specific box. This reveals the "bias" or the hidden rules the AI has learned from its training data.

3. The Performance: Humans vs. Robots

Now comes the magic part. Two "actors" take the stage:

A Human Performer: They watch a video generated by the AI that shows how to move, then they try to copy it.
A Quadruped Robot: A four-legged robot (like a high-tech dog) that receives the exact same instructions but translates them into mechanical code.

They perform on a mirrored floor. The mirror is important because it makes the human and the robot look like they are in the same world, reflecting each other.

The Analogy: Imagine a human and a robot playing a game of "Mirror, Mirror." The human is trying to be fluid and emotional, while the robot is trying to be precise and mechanical. But both are following the same script written by the AI. The tension comes from watching the human struggle to be "free" while the robot executes the commands perfectly.

4. The Archive: Turning Movement into Data

As they dance, cameras track every single move. The system turns their physical movements into 3D digital trails (like glowing lines in the air that show where they moved).

The Analogy: This is like leaving footprints in digital sand. Even after the dance is over, the "footprints" remain. The researchers save these digital footprints to study later. They call this an "Archaeology of AI" because they are studying the "fossils" of how AI interprets human emotion and movement.

Why Does This Matter?

The project isn't just about making cool robot dances. It's a critique.

The Big Question: If an AI is trained on millions of books and movies, does it learn that humans are supposed to be strong, or weak? Does it learn that women should move one way and men another?
The Revelation: By forcing a human and a robot to follow the AI's strict rules, ReTracing shows us the invisible cage the AI builds. It reveals that AI isn't just a neutral tool; it has its own "personality" and biases that dictate how we move, think, and exist.

In a nutshell:
ReTracing is like a detective story where the clues are dance moves. It uses a robot and a human to expose the hidden rules of Artificial Intelligence, showing us that when we let AI choreograph our lives, we might be dancing to a tune we didn't choose.

1. Problem Statement

The paper addresses a critical gap in the intersection of generative AI, robotics, and performance art: the lack of critical frameworks to examine how AI systems encode socio-cultural biases, control mechanisms, and disciplinary power into physical movement.

The Gap: While recent advances in diffusion models allow for text-to-motion generation and robotics can execute precise movements, these technologies are often treated as culturally neutral tools. Existing work rarely contrasts human and robotic agents to reveal how generative systems interpret "embodied identity" (race, gender, emotion) differently across biological and mechanical substrates.
The Core Question: How do generative systems shape, constrain, and produce bodily movement, and what does it mean to be human among AIs that also move, think, and leave traces?

2. Methodology

ReTracing employs a multi-agent, embodied performance framework that functions as a form of "digital archaeology." The workflow consists of four distinct stages:

A. Data Excavation (Literary Input)

The system uses excerpts from seven science-fiction and dystopian novels (Frankenstein, The Yellow Wallpaper, The Handmaid's Tale, Dawn, Poor Things, Parable of the Sower, Klara and the Sun) that depict human-machine interactions, fear, and control.

LLM Processing: The team utilizes Qwen-2.5 (temperature 0.7) to analyze these excerpts.
Prompt Engineering: For each excerpt, the LLM generates two types of instructions:
1. Positive Prompts (Activations): Descriptions of movements the agent should perform.
2. Negative Prompts (Restrictions): Descriptions of movements the agent must not perform.
- Goal: To categorize movement based on algorithmic logic of "permitted" vs. "prohibited" actions, mirroring the constraints found in the source texts.

B. Multi-Agent Choreography Generation

The generated prompts are translated into executable actions for two distinct agents:

Human Performer: A text-to-video diffusion model (specifically Wan2.1) converts the prompts into short video clips. These videos serve as visual choreographic guides for the human, simulating motion based on the AI's interpretation of the literary affect.
Quadruped Robot (Unitree Go2): The same prompts are converted into sequences of predefined robotic actions (e.g., "Pounce," "Shake Hands," "Move Forward"). The robot executes these with mechanical precision using its own programmed language.

C. Performance and Capture

Environment: The performance takes place on a mirrored floor, creating a space for introspection where the body is both subject and object.
Capture: Multi-camera setups record the simultaneous enactment by the human and the robot.
Concept: The mirror and the shared visual plane highlight the tension between organic fluidity and mechanical rigidity.

D. Reconstruction and Archiving (Digital Traces)

3D Reconstruction: Monocular video footage is processed using a monocular 3D point tracking model (SpatialTrackerV2).
Output: The system infers joint positions over time to produce temporally consistent 3D skeletal keypoints and motion trails.
Archive: These 3D point clouds form a "digital archive" of motion traces, converting the ephemeral performance into traceable data forms.

3. Key Contributions

Archaeology of AI Framework: Introduces a novel methodology to "excavate" the logics embedded within generative systems by tracing the pipeline from literary text $\to$ AI prompt $\to$ embodied movement $\to$ digital trace.
Comparative Agent Analysis: Unlike prior work focusing solely on human agents, ReTracing explicitly contrasts human and robotic agents. This reveals how the same algorithmic logic (prompts) results in divergent physical expressions based on the substrate (biological vs. mechanical).
Bias and Control Visualization: The project demonstrates how generative models encode socio-cultural biases (e.g., gendered movements, restrictions on emotion) into choreography. It visualizes "disciplinary power" where the AI dictates what bodies can and cannot do.
Open Dataset: The authors commit to releasing the complete 3D motion trace dataset and the technical workflow, enabling future research into AI-driven choreography and bias detection.

4. Results

Execution: The system successfully generated distinct choreographies for both agents based on the same literary inputs.
- Example: In The Handmaid's Tale, the human performed "small steps" and "trembling," while the robot executed "Stand Up from Fall" and "Stretch," highlighting the different ways "subservience" or "restriction" is interpreted by AI for different bodies.
Visualization: The 3D reconstruction successfully captured the temporal evolution of movements, creating visual "trails" that serve as a data archive of the performance.
Segmentation: The application of the Segment Anything Model (SAM) to the interaction scenes successfully distinguished between interacting agents and the environment, validating the technical pipeline for downstream motion analysis.

5. Significance and Implications

Theoretical Impact: ReTracing reframes generative AI not just as a computational tool, but as a multi-agent system of inscription. It posits that AI is as much about control and memory as it is about computation. By treating AI as an archaeologist of the present, the work reveals how algorithms inscribe values onto bodies.
Critical Reflection: The project challenges the notion of the "universal body" in motion synthesis. It exposes how AI reinforces stereotypes (e.g., feminized bodies, restricted movement) through its training data and prompt logic.
Ethical Considerations: The paper acknowledges the risks of such systems, including the potential for reinforcing biases and privacy concerns regarding the collection of body shape and motion data. It calls for careful contextual framing to ensure the work is interpreted as a critique of algorithmic bias rather than mere entertainment.
Future Research: By opening the dataset, ReTracing invites the community to explore how different prompts alter the "choreography of control" across diverse agents, fostering a deeper understanding of the relationship between language, code, and physical movement.