Ropedia Academy

🌐 Live site · 4 tracks · 36 bilingual lessons · live 3D demos · spaced repetition

An interactive, bilingual (中文 / English) course on embodied & spatial AI — four connected tracks covering human modeling & motion, 3D/4D reconstruction & neural rendering, egocentric vision & interaction, and scene reconstruction & world models. Read lessons, play with live interactive demos (including real-time 3D), run the code in Colab, self-test, and review with spaced repetition. Runs entirely in the browser — no account required.

Features

• 4 tracks · 36 lessons, each with a bilingual explanation, key terms, key papers, external links, cross-track links, and self-check questions.
• An interactive demo in every lesson — real-time three.js 3D (Gaussian splatting, a raymarched NeRF volume, an articulated SMPL body) plus explorable diagrams (triangulation, bundle adjustment, rotation continuity, hash grids, TSDF fusion, SLAM loop closure, reference frames, world-model rollouts, …).
• Math & code — KaTeX formulas and a runnable Python/PyTorch snippet per lesson, each with one-click Open in Colab; its real output (printed values
  • figure) shows inline behind a predict-it, then reveal toggle.
• Bilingual reading (中文 / English / 双语) and hover-to-define glossary tooltips on foundational terms.
• Self-graded checks & quiz mode, spaced repetition (SM-2) with a 7-day forecast, a cross-track concept map, progress tracking, and a ⌘K palette.
• Light / dark theme, mobile-friendly, local-first (no account required).

Training labs

Beyond the per-lesson snippets, notebooks/training/ holds twenty-three real, multi-cell Colab notebooks you can actually train — split into clear blocks (data · model · train · compare) so you can step through and watch each stage:

The twenty self-contained PyTorch labs are verified to train (each was run to confirm loss drops / PSNR climbs / metrics beat chance); the three foundation labs follow the official APIs and run on a Colab GPU. Every lab records its checkpoint + loss/eval history + figures to a downloadable outputs/<lab>/, and an optional cell publishes the run to the Hugging Face Hub (a model repo with a metrics-and-plot model card) so you can gather them into a Collection. Open them from the dashboard's Training labs section, or see notebooks/training/README.md for one-click Colab badges. Set Runtime → T4 GPU first.

Advanced labs (heavy · real repos · GPU)

notebooks/advanced/ adds twenty-two heavy GPU pipelines on real research repos for when you want the production tools, not a teaching toy:

The verified self-contained labs were each run to confirm real results; the advanced labs follow each project's official recipe and run on a Colab GPU (some need gated data), so they are flagged as not pre-executed. Every lab records its checkpoints / loss-eval history / outputs to a downloadable folder, every lab cross-links to its related tracks (chips + filters + an in-notebook note), and the Labs tab closes with a Future explore directions board — open cross-track projects that combine the labs.

These clone official repos, download multi-GB checkpoints/datasets, and require a GPU — they're authored to each project's documented recipe and are not pre-executed (expect to pin a version or two). See the dashboard's Advanced labs · GPU section or notebooks/advanced/README.md.

Models on Hugging Face

Every lab publishes to the Hub as a model repo with a full card (base model · objective · dataset · config · evaluation · inference · limitations · failure cases · license · citation · reproducibility), gathered into one Collection. Try them all live in a single Space.

The 🚀 demos Space runs real models per task: a small instruct LLM (chat · next-action · ReAct tool use), a class-conditional DDPM that generates digits, a CartPole agent, and a world-model CEM planner — plus a gallery of every trained model's figure + metrics.

19 trained · 26 documented placeholders · 45 repos total — one click trains a placeholder into a real model.

A · Human modeling & motion

B · 3D / 4D & neural rendering

C · Egocentric vision

D · Scene & world models

LM · Language & multimodal

AG · Agents & RL

Run & deploy

npm install
npm run dev      # http://localhost:5173
npm run build    # production build → dist/

Pushing to main auto-builds and deploys to GitHub Pages via the included GitHub Actions workflow (.github/workflows/deploy.yml) — that's what serves the live site above. It's a static SPA, so dist/ can also be hosted on any static host (Netlify, Vercel, Cloudflare Pages).