Push Anything

Project website for Push Anything: Single- and Multi-Object Pushing From First Sight with Contact-Implicit MPC, hosted at https://dairlab.github.io/push-anything/.

The site is static HTML, CSS, and JavaScript. GitHub Pages serves it directly; no package installation or build step is needed.

Media loading

• The opening and credits videos autoplay silently in loops without controls. They pause offscreen or in background tabs, and resume when visible again. The credits video source loads only when its section comes into view.
• Overview and experiment videos have no src until a visitor presses play. Selecting a single-object grid thumbnail also starts its clip.
• A shared experiment player exposes all 61 original demo choices, filtered by object count. Two-, three-, and four-object experiments use object-combination chips with recording counts. Previous/next arrows browse recordings within a combination; the counter shows the current video. Choosing a combination resets to its first recording. Browsing combinations or recordings updates the poster without requesting a video.
• The original five-by-five object grid and 25 rotating GIF previews are preserved, including Push-T. The Push-T player poster uses a still frame from its hardware video. Small WebP stills appear first; animated GIFs are requested only when their choices approach the viewport. Visitors who prefer reduced motion keep the still previews. Other previews and figures load lazily.
• Changing object categories cancels the current experiment request without downloading another clip. Only one user-selected video plays at a time, and offscreen/background videos pause.
• Intro and overview MP4s have their metadata moved to the front for faster startup, with no re-encoding or quality loss.
• System fonts and native browser controls avoid font, library, and icon CDN requests.

The original long experiment clips retain their MP4 layout and size. To inspect remaining clips, run:

python3 tools/optimize_videos.py

To losslessly remux them after installing ffmpeg:

python3 tools/optimize_videos.py --apply

The default command only inspects files. --apply replaces each clip atomically after successful conversion and a metadata-layout check. Rewriting the remaining clips changes about 3.6 GB of tracked video data, increasing Git history; review storage and GitHub publishing limits before committing that separate change. For a much larger collection, external video hosting or adaptive streaming can reduce the repository footprint and improve delivery.

Local preview and checks

python3 -m http.server 8000
python3 -m unittest discover -s tests -v
node --check script.js

Open http://localhost:8000/ to preview. The basic Python server is sufficient for layout checks; use a server supporting HTTP byte ranges for realistic seeking/playback tests.