Media catalog and storage#

Keep selected images, diagrams and captions in this repository. Keep large videos in versioned release assets, and original recordings in backed-up lab storage. The goal is for a future maintainer to understand an asset, locate its source and replace its host without rewriting the explanation around it.

The selected stacking dataset and calibration captures use separate Google Drive downloads, listed on the dataset page. They are not embedded video assets. Keep a versioned README and checksums with each download, and verify access without the owner’s signed-in account before adding its link to the guide.

Included figures#

The following original PNGs were visually inspected and copied unchanged from g1-aprilcube-demo; identical copies are retained in the consolidated main at 2b7274b11f1862ebfcd05b48ff678d995e55269e:

Stable ID	Figure	What it shows
`dex3-descriptor-states`	Descriptor audit	Rendered right/left open, half and closed states with learned descriptor proxy boxes
`u-support-orientations`	U supports	Six geometric support orientations; printed counts are geometric candidates
`assembly-scene`	Assembly scene	A rendered G1/table/part planning scene, not a photograph

Two additional CAD views are copied unchanged from the fixture repository at f295def18bd936031fba325d4e8bdfc71fea671a:

Stable ID	Figure	What it shows
`dex3-dorsal-mount-cad`	Wrist-marker placement	Right-hand plate, screw arrangement and support pads; a CAD view, not a physical fit measurement
`g1-torso-v7-cad`	V7 torso structure	Modeled crossbars, side plates and board; the assembly still has unresolved fit checks

The machine-readable catalog is docs/assets/media.json. It records exact source paths, commit, SHA-256, context, review status and caption. The source PNG labels describe their original experiments; the surrounding chapter supplies later corrections and scope.

Photographs and installation figures#

Selected photographs from the author’s September collection are included in Git. Their framing, dimensions and decoded pixels are unchanged; embedded location and other EXIF/text metadata were removed from the publication copies. The originals and the PowerPoint remain separate source records.

Stable ID	Where it is used	Source
`realsense-pc2-usb-ports`	Camera troubleshooting	August 3 phone photograph
`realsense-head-pitch-witness-mark`	Head-pitch reference	September 21 photograph; author identifies the white witness mark
`g1-wrist-spacing-caliper`	Wrist-spacing investigation	Author-supplied original photograph; caliper display reads 51.38 mm; the chapter distinguishes that reading from verified joint-axis spacing
`printed-target-collection`	Printed targets	September 21 photograph of cubes, wrist carrier and board
`cushioned-chair-reference`	Physical support	Chair with its cushion; September 21 photograph
`rigid-chair-reference`	Physical support	Same chair with the cushion removed; author-supplied 850 × 1133 px photograph
`chair-support-comparison`	Physical support	Both original chair photographs, combined side by side in one figure
`dex3-connector-routing`	Hand fitting	`Dex3 Hands.pptx`, slide 7, `ppt/media/image11.png`
`dex3-hand-seated`	Hand fitting	Same slide, `ppt/media/image10.png`
`dex3-cable-routing`	Cable routing	Slides 4 and 8, `ppt/media/image2.jpg`
`dex3-wrist-fasteners`	Hand fitting	Slide 6, `ppt/media/image5.jpg`

The installation figures now retain the features indicated by the slide annotations, with overview photographs, enlarged details and numbered callouts. They identify hardware and the recorded fitting sequence; the electrical pinout and complete power-isolation procedure still need author review.

Annotated hardware figures#

Use an overview to locate a feature and a close-up when the detail would be too small at normal page width. Keep each caption and instruction in Markdown below the figure. The hardware pages use editable SVG overlays, so labels, circles and leader arrows stay separate from the original photo pixels.

The layouts are in docs/assets/hardware-figures.json. Each records its source image hashes, panel crops, feature positions and source photograph or slide. The Dex3 features were checked against rendered slides 6–8. Their presentation adapts the slide callouts into rings and numbered detail panels; it is not a screenshot of the whole slide. The head-pitch mark comes from the author’s separately supplied reference photograph.

The chair comparison uses the same layout system to place two cropped photographs side by side in a single image, with one Markdown caption. The author confirms they show the same chair with and without its cushion. Both complete chairs remain visible; the crops remove excess background without stretching either photograph.

To adjust a figure, edit its layout and regenerate it:

python3 tools/build_hardware_figures.py
python3 tools/build_hardware_figures.py --check

This standard-library helper uses the checked-in publication photographs; neither PowerPoint nor LibreOffice is required. It verifies their hashes before rendering. SVGs embed those photos, so a downloaded SVG is self-contained. Review feature placement, magnification, alt text and the Markdown caption on desktop after a change. The CI check catches stale generated figures.

Embedded demonstrations#

The September media release hosts the publication videos. Each player has native controls, a local poster, a descriptive caption and a separate download link. Videos load on request and do not autoplay. All publication copies are silent; the original seat, physical-stacking and calibration movies retain audio that has not been reviewed here.

Stable ID	Chapter	Evidence shown
`g1-physical-cube-stacking`	Pickup and stacking	Physical pickup, stack, release and hand withdrawal; no software handback display
`g1-standing-calibration`	Calibration capture	Physical pose sequence with wrist markers; no fit or cleanup result displayed
`dex3-tactile-rviz-demo`	Tactile sensing	Changing taxel visualization; not calibrated force
`dex3-simulated-retention-demo`	Grasp qualification	Rendered Isaac/PhysX retention replay
`curobo-stack-plan-demo`	Stacking plan	Rendered geometric phases through placement
`mujoco-rviz-short-demo`	MuJoCo	Side-by-side RViz and simulation
`g1-cushion-seat-setup`	Physical support	Seated robot, harness, targets and arm movement
`g1-lerobot-viewer-walkthrough`	Dataset inspection	Images, traces, 3D replay and diagnostics

The release’s checksum file identifies the hosted copies. The catalog records each original hash separately from the publication hash, along with the encoding recipe and poster timestamp. The full visual duration is retained; the physical videos are compressed for web playback, not sped up. The viewer’s odd source height is padded by one row for H.264 encoding.

Asset	Duration / dimensions
Arm demo	34.28 s / 1856 × 1044
Dex3 demo	18.49 s / 1824 × 1048

Author-supplied September media#

The collection now contains a nine-slide Dex3 installation deck, five photographs and eight videos. The originals remain outside this repository. The public media catalog retains original/publication hashes, selected source paths, captions and review limits. Detailed working reviews remain private.

Photographs, embedded deck images and sampled video frames were inspected. Audio in the three physical recordings remains unreviewed. The short MuJoCo clip is a separate file from the July release assets reviewed above. The deck’s linked boot-calibration movie was not present in the folder.

Recreating a publication copy#

The ordinary site build needs only this repository. To prepare media again, obtain the matching original collection and install ffmpeg/ffprobe, then run the standard-library Python helper from the repository root:

python3 tools/prepare_media.py \
  --source-dir /path/to/originals \
  --video-dir .local/publication-media \
  --asset g1-physical-cube-stacking

Omit --asset to prepare all assets that have a preparation recipe in the catalog. Unchanged source-repository figures are copied from their pinned source paths; generated SVG layouts use the separate figure helper above. The preparation helper checks source hashes, writes selected images/posters under docs/assets/, puts video outputs in the specified directory and updates their catalog hashes. Review the resulting diff and playback before uploading. Encoder versions can change file bytes, so record the new output hash rather than expecting byte-identical transcodes. Use a new versioned release when changing a published video; retain the old asset for existing references. The helper never uploads or changes originals.

Adding a new asset#

Choose a stable descriptive ID, such as dex3-extension-routing.
Record the run/date, hardware or simulation context, creator and exact source.
Preserve the original in the lab archive. Add a publication copy and its checksum to the catalog; keep an accessible caption and meaningful alt text.
Explain what the asset establishes, including the phase shown and any outcome that happens outside the clip.
Check the rendered page on desktop during drafting; mobile presentation is deferred to the final review. For video, test playback, controls and the download fallback separately.

Use ordinary Markdown images and simple HTML video only when needed. Avoid autoplay, credentials and expiring signed URLs. Stable, intentionally shared Drive download links are appropriate for the datasets above. An unedited failure clip can be more informative than a polished success montage when its caption explains the first failure and the later correction.

Ownership during the lab handoff#

GitHub Release links are a practical starting host, not the original archive. The lab copy must either retain access to those assets or copy them to a lab-owned release and update catalog URLs. A repository copy alone does not copy another repository’s release assets.

Keep raw bags and full media collections outside the documentation build. Before changing storage, check the host’s current file-size, Pages and LFS rules: large files, Pages limits, LFS restrictions.