BONES-SEED Viewer

A web application for browsing the BONES-SEED motion capture dataset. Renders skeleton animations on 3D character models in the browser using Three.js, with a searchable file browser, metadata panel, and temporal label overlay.

Supports two character models:

• SOMA body model
• Unitree G1 robot

Temporal labels were created by NVIDIA for the Kimodo project.

Quick Start

Requires Docker (includes Docker Compose).

git clone <repo-url>
cd seed-viewer

cp .env.example .env
# Edit .env — set DATA_PATH to your extracted BONES-SEED dataset root

./shdocker.sh          # Linux / macOS
# or
docker compose up --build   # any platform

The app will be available at http://localhost:8666.

Dataset Structure

DATA_PATH should point to the root of the extracted dataset:

DATA_PATH/
  metadata/               # Parquet metadata + jsonl temporal labels
  soma_proportional/bvh/  # Original mocap on SOMA — BVH files
  soma_uniform/bvh/       # Mocap retargeted to unified SOMA shape — BVH files
  g1/csv/                 # Mocap retargeted to G1 robot — MuJoCo compatible CSV files
  soma_shapes/			      # SOMA shape parameters

Local Development

For development without Docker:

• Node.js 20.10.0 (see .nvmrc)
• Python 3.11
• PDM for Python dependencies

Backend (terminal 1):

cd backend
pdm install
DATA_ROOT=/path/to/dataset PORT=8080 python src/main.py

Frontend (terminal 2):

cd frontend
npm install
npm run dev    # Vite dev server on localhost:5173, proxies /api → localhost:8080

Testing

End-to-end tests use Playwright against the running app at http://localhost:8666:

cd frontend
npx playwright install   # first time only
npx playwright test      # headless Chromium
npx playwright test --headed   # with visible browser

SOMA Example

The soma/ directory contains a minimal Python example for parsing BONES-SEED motion capture data and running it through the SOMA body model.

Related Work

BONES-SEED is part of a larger effort to enable humanoid motion data for robotics, physical AI, and other applications.

Check out these related works:

• SOMA Body Model - Parametric human body model with standardized skeleton, mesh, and shape parameters
• SOMA Retargeter
• GEM-X - Human motion estimation from video
• Kimodo - Kinematic motion diffusion model for text and constraint-driven 3D human and robot motion generation
• ProtoMotions - GPU-accelerated simulation and learning framework for training physically simulated digital humans and humanoid robots
• SONIC - Whole-body control for humanoid robots, training locomotion and interaction policies

License

Apache 2.0 — see LICENSE.