This project is a responsive static research-team website for Darwin Agent.
DA.pptx— the Darwin Agent keynote deckmimem.pdf— the Mi-Memory technical report and related memory notes
The page positions the team around continuously learning, self-evolving agents across digital and physical worlds and links to the public Darwin-Agent GitHub organization. It keeps the original editorial layout, with new-project shortcuts, research directions, and project-specific resource links.
Official team site: https://darwin-agent.github.io/
Maintained in Darwin-Agent/darwin-agent.github.io. This organization-owned deployment preserves the original website and logo work by Shaokun in shaokun91/darwin-agent-website, including its Git history. The personal repository is the source provenance, not the canonical team website. No new license is assigned to the original assets.
index.html— research homepagestyles.css— visual design and layoutscript.js— responsive navigation and interactive evolution visualizationassets/darwin-agent-qr-da.svg— self-contained DA-branded homepage QR for the footerassets/darwin-agent-qr-da.png— 1176 × 1176 downloadable QRassets/darwin-mark-da-orange.png— 1024 × 1024 transparent DA monogramassets/darwin-mark-da-white.png— matching white transparent DA monogramassets/darwin-agent-banner.png— shared organization profile and social-preview bannerscripts/build_qr.py— deterministic QR asset generator (development only)
- HarnessPAI — an evolving executable harness for Physical AI; paper · code
- WIA — world modeling and action learning across seven digital environments; code · demos
- HarnessX — composable and evolvable agent harnesses
- Mi-Memory — composable, evolvable multi-source memory
- CAR-bench TRACE — trajectory-driven skill-bank evolution
- Topological Collapse — higher-order structure in agent societies
- Digital Agents Meet World Models — survey and research collection
Open index.html directly in a browser, or run a local server:
cd darwin-agent.github.io
python3 -m http.server 8000Then visit:
http://localhost:8000
- Experience → Action → Evolution as a system-level research ambition
- A dedicated
#learningsection introducing the research approach - Three stages connecting accumulated experience, action, and evaluated feedback
- Harness as a first-class, evolvable behavior object
- Memory as a lifecycle framework with provenance and governance
- World modeling and action learning across digital environments
- Physical AI through programs that stay fixed within a rollout and evolve between rollouts
- Multi-agent systems as a topology problem
Run the dependency-free content and link-structure checks:
python3 -m unittest discover -s tests -vBefore publishing, also check desktop and mobile layouts, keyboard navigation, reduced-motion behavior, and new external resource links in a browser. The static checks do not make network requests or replace browser QA.
GitHub Pages publishes the organization repository root from main; no build step or client
dependencies are required. Push reviewed changes to main, wait for the Pages
deployment, and verify the live site.
Canonical metadata, social sharing, robots.txt, sitemap.xml, and all homepage
QR assets use https://darwin-agent.github.io/. The organization profile and its
website field link to this address. Updating or redirecting the original personal
deployment requires its owner's permission and is not part of this repository.
HarnessPAI and WIA descriptions are grounded in their official project pages, checked on 2026-11-12. The homepage summarizes their mechanisms and scope rather than reproducing benchmark rankings or implying that every advertised artifact is bundled with the public code. Project resources should link to verified public destinations.
Dates were checked on 2026-11-12. First preprint means the initial arXiv
submission (v1), not a later revision, conference acceptance, or repository
creation date. Report published follows an explicit dated announcement from
the project. Keep the source's date precision: do not invent a day when only a
month is documented. Each displayed date uses a semantic time element; its
parent's data-date-source records the evidence URL without nesting links inside
clickable project cards.
| Project | First preprint / report date | Primary evidence |
|---|---|---|
| HarnessX | 2026-06-12 | arXiv v1 |
| HarnessPAI | 2026-09-24 | arXiv v1 |
| Mi-Memory | 2026-07-21 | arXiv v1 |
| CAR-bench TRACE | 2026-08-24 | arXiv v1 |
| Topological Collapse | 2026-08-16 | arXiv v1 |
| Digital Agents Meet World Models | 2026-05 | Official report announcement; distinct from the April repository launch |
| WIA | Not displayed; no verified first-publication date | Official project page and repository do not currently state one |
The hero's evolve. highlight uses Xiaomi orange (#ff6900), matching the
HarnessPAI project page. Other existing color accents are retained.
The page ends with a scan-and-share section linking directly to
https://darwin-agent.github.io/. There is no redirect service,
tracking URL, third-party image dependency, or runtime QR library. The center
uses the reviewed image-generated DA monogram, normalized to exact Xiaomi
orange (#ff6900) on white. The same transparent mark appears in the header,
footer, and favicon. The previous D-only assets remain available at their old
paths; no existing downloadable asset was deleted.
The QR explicitly retains version 6, H error correction, a four-module white quiet zone, and a small central logo plate clear of finder, timing, and alignment patterns. The SVG embeds its logo and is independently portable; the PNG download is 1176 px.
To rebuild, install qrcode==8.2 and Pillow in a development virtual environment
and run python scripts/build_qr.py. To run image-decoding tests as well, install
zxing-cpp==2.3.0 and run python -m unittest discover -s tests -v. The tests check
the decoded URL at several sizes, after rotation, JPEG compression, and grayscale
conversion. Recheck the decoded URL and layout whenever the destination or logo
changes. The generator intentionally rejects a different QR version so logo
placement cannot silently cover a function pattern.
The legacy darwin-agent-qr.svg and darwin-agent-qr.png paths are retained and
rebuilt to the same official team URL, with their original D-only logo.
Use the orange transparent PNG on light or dark neutral backgrounds and the white variant where a reversed mark is needed. Preserve the clear space baked into the asset, do not stretch it, and do not add the old square border. Use the monogram alone at small sizes; pair it with the existing Darwin Agent wordmark for larger website and presentation applications. A 40 px mark is used in the website navigation; the QR logo remains within its tested central safe area.
The homepage uses the mark in navigation, the footer, and the QR code. Public
copy focuses on the team's research rather than explaining the letter geometry.
The #learning section introduces the three-stage learning loop as a research
goal, not a claim that all public projects already implement an autonomous,
end-to-end evolution loop.
The current asset uses the reviewed softer third candidate. Its silhouette is uniformly scaled and centered to match the previous mark's optical size without stretching. The orange and white versions share the same alpha channel; the organization banner and QR code are rebuilt from the orange version. The earlier assets remain recoverable through repository history.