Project page for "Mask2Real-WM: Segmentation Masks as a Sim-to-Real Bridge for Controllable Dexterous World Models" (CoRL 2025).
A two-stage action-conditioned world model for dexterous manipulation that decouples dynamics (segmentation-mask prediction, pretrained on >50 h of simulation) from rendering (RGB, trained on <2.5 h of real data), using segmentation space as a sim-to-real bridge for 23-DoF control.
🌐 Live site: https://<your-username>.github.io/<repo-name>/
index.html # Project page (abstract, method, results, citation)
paper.pdf # Full paper
demo/ # Self-contained interactive DOF-controllability demo (Three.js)
index.html # 3D ORCA hand viewer + per-DOF prediction videos
scene.json # Mesh transforms / DOF mapping
meshes/ # STL meshes for the hand + arena
videos/ # Per-DOF prediction videos for each model & sample
assets/
img/ # Figures (teaser, method, results, qualitative)
video/ # Long-horizon rollout comparison videos
.nojekyll # Tells GitHub Pages to serve files as-is (don't run Jekyll)
Embedded in the main page (and openable full-screen at demo/index.html), the demo lets
you click any of the 23 degrees of freedom on a 3D model of the ORCA hand and watch the
world model's prediction when that single action component is perturbed by a sinusoid.
You can compare four models — WM + LoRA (ours), WM Mid-train, WM Real-Only,
and the monolithic Baseline — across multiple evaluation samples.
- Create a new, empty GitHub repository.
- From inside this folder:
git init git add . git commit -m "Add Mask2Real-WM project website" git branch -M main git remote add origin git@github.com:<your-username>/<repo-name>.git git push -u origin main
- In the repo: Settings → Pages → Build and deployment → Source: Deploy from a branch,
pick
main/root, and save. The site goes live at the URL above in a minute or two.
Everything is static — no build step. The
.nojekyllfile ensures thedemo/folder and all assets are served verbatim.
- Update the Code link in
index.html(id="code-link") once the code repo is public. - The 3D demo loads Three.js from a CDN, so it needs an internet connection to render the hand.
- Total size is ~95 MB (mostly demo meshes and per-DOF videos); all files are well under GitHub's 100 MB per-file limit.