Skip to content

About

RoboChrono: A Real Robot Benchmark for Streaming Task Understanding

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

31 Commits

Folders and files

Repository files navigation

RoboChrono project page

A static research website for RoboChrono: A Real Robot Benchmark for Streaming Task Understanding.

This edition builds on the original li-zzmm/robochrono project page and its recordings. The public website source is maintained in continuity3/robochrono. The website’s GitHub entries link to the benchmark repository, mfan-res/ROBOCHRONO. The public preprint is hosted in the benchmark repository: read the paper.

Preview locally

From the repository directory, run:

python3 -m http.server 8765 --bind 127.0.0.1

Then open http://127.0.0.1:8765. No installation or build step is required. The site also supports opening index.html directly; clipboard access may require the local server or manual copying.

Content and assets

  • index.html: paper summary, authors, affiliations, capabilities, findings, and citation.
  • styles.css: white visual theme, responsive layout, and reduced-motion styles.
  • script.js: Tianji/GIM hero preview switching, embodiment tabs, continuous video gallery, visibility-aware videos, motion controls, video/figure dialogs, and citation copying.
  • assets/brands/: supplied Tianji, GIM, and Shenzhen University marks, displayed in that order. The Tianji SVG frames an embedded, unchanged original image; the Shenzhen University SVG is extracted from the supplied Illustrator artwork.
  • assets/figures/: figures from the public preprint and supplied figure PDFs. The overview and capability chart use the final public-version figures, with provenance and companion results data. SVG text is outlined so it renders without external fonts.
  • assets/paper/robochrono.pdf: a matching local copy of the public preprint. The three visible Paper entries use the canonical GitHub PDF link.
  • assets/videos/: original repository recordings; Stack Cubes previews are already encoded at 2× speed. The Tianji hero excerpt uses seconds 16–34 of stack-cubes-tianji-2x.mp4, preserving its 2× speed. The hero defaults to Tianji and can switch to the existing 20-second stack-cubes-gim-main-2x.mp4 recording using the thumbnail tabs below the video.
  • assets/images/gim-hero.jpg: GIM preview thumbnail extracted at 6 seconds from the original Stack Cubes recording.
  • assets/videos/gallery/: trimmed, silent web previews from yyyyywv/egocentric, with posters and a per-clip source manifest. Gallery excerpts play at their original speed and are attributed under CC BY-NC 4.0.

The author list follows the supplied author screenshot, with Shihao Li placed after Yiyang Ma. The first seven authors are marked as equal contributors; Zhuo Xu, Long Chen, and Ruoxiang Li are marked as corresponding authors. Affiliations incorporate the author-requested updates: Yuzhou Wu lists Tianji Tec. and Shenzhen University; Longteng Fan, Shihao Li, Yifan Wu, Zichen Zhang, Ruiqi Yang, Weibin Kong, Yihang Xu, Haoran Liu, Tao Xu, Zhuo Xu, and Long Chen list General Intelligence Machine only.

Displayed results are scoped to the manuscript's reported evaluation. Model coverage and aggregation differ; the visual-evidence ablation uses five open-weight models on a balanced subset of 312 stems.

GitHub Pages

The publication target is continuity3.github.io/robochrono. In the repository Pages settings, select Deploy from a branch, then main and / (root). The root contains .nojekyll and uses repository-relative asset paths, so no build step is required. The absolute social-image and citation URLs in index.html use this target.

The displayed model-capabilities-public.svg comes from the public preprint’s corrected Figure 4 and matches all 18 models / 180 reported values in its Table 3. Its companion JSON preserves the reported group scores. benchmark-overview-public.png uses the public preprint’s causal-evaluation overview; earlier source figures remain archived in the assets.

The dataset entries list Tianji Dataset (gimai/RC-Tianji) first, followed by GIM Dataset (gimai/RC-GIM), in both the hero and resources sections. GitHub entries link to the benchmark repository, mfan-res/ROBOCHRONO. All three Paper entries open the public preprint in the benchmark repository. The public website source repository remains available separately.

Search discovery

The page declares https://continuity3.github.io/robochrono/ as its canonical URL. sitemap.xml lists that same URL; section fragments such as #in-motion are anchors within this one page, not separate pages to submit.

To inspect Google's actual indexing status, use a verified URL-prefix property for https://continuity3.github.io/robochrono/ in Google Search Console. Inspect the homepage, run the live URL test, and request indexing if eligible. Submit https://continuity3.github.io/robochrono/sitemap.xml in the Sitemaps report. A sitemap helps discovery but does not guarantee indexing or rankings.

The effective robots policy is at the hostname root, https://continuity3.github.io/robots.txt, which belongs to the separate personal-site repository. A robots.txt inside /robochrono/ would not control crawling. This project does not alter the personal site's crawler policy or sitemap.

Browser checks

Verified in Chrome at 320, 390, 768, 1024, and 1440 px widths, including keyboard tab switching, image dialogs, citation copying, video playback controls, reduced motion, and readable content without JavaScript.

About

RoboChrono: A Real Robot Benchmark for Streaming Task Understanding

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages