icra-supplementary

ICRA Supplementary

At most ML conferences "supplementary" means an unlimited appendix PDF. ICRA has

no such thing: the 2026-cycle rules cap the reviewed document at 6 content pages

plus 2 reference-only pages, with no appendix upload. What ICRA has instead is

the video attachment — a robotics-native evidence format with hard specs

(≤ 180 seconds, ≤ 20 MB, mpeg/mp4/mpg in 2026) and its own submission windows.

This skill treats the video as the serious scientific artifact it is.

The self-containment rule

Because reviewers are not obligated to open anything beyond the PDF, and because

there is no appendix:

• Every load-bearing definition, equation, protocol detail, and result **must be

in the 6 pages**. The video may demonstrate; it must never be the only place a

claim is supported.

• External links (project page, code) are pointers to optional corroboration.

Under the double-anonymous policy they must also be anonymized or omitted.

• The test: delete the video and all links mentally — does the paper still prove

its claims? If no, restructure the body, not the supplement.

What the video is for

Robotics reviewers watch the attachment early, often before reading past the

abstract. It answers questions text cannot:

| Reviewer question | Video answer |

|---|---|

| Is this real hardware or renders? | continuous, uncut executions |

| Is the demo cherry-picked? | multiple trials, varied conditions, a counter on screen |

| How fast is it really? | wall-clock timestamps, explicit "1× speed" labels |

| What does failure look like? | one or two failure clips with a caption naming the cause |

| Is a human secretly helping? | wide shots showing the full workspace and no puppeteer |

Storyboard for 180 seconds

0:00-0:10  title card: paper title only (no authors — double-anonymous)
0:10-0:30  task + platform in one wide continuous shot
0:30-1:30  core capability: 3-5 uncut executions, varied instances,
           on-screen labels: condition, trial number, 1× speed
1:30-2:10  the money comparison: baseline vs method, same task instance,
           side by side if possible
2:10-2:35  a failure case, honestly labeled with the taxonomy category
           from the paper's failure table
2:35-3:00  generalization montage (may be sped up — label "4×"),
           closing card pointing to the paper, nothing else

Rules the storyboard encodes: label every speed change (silent speed-ups are

treated as deception when caught); prefer uncut takes over montage for the core

claim; synchronize the video's trial conditions with the paper's tables so a

reviewer can cross-reference; show failure deliberately — it converts the video

from advertisement into evidence.

Meeting 20 MB without looking like 2005

180 seconds under 20 MB means roughly 900 kbps total — tight but workable:

• Encode H.264 (in an mp4 container per the allowed formats), 720p, high-profile

two-pass; drop the audio track entirely (narration goes into captions, which

also survive muted review viewing and avoid voice de-anonymization issues).

• Avoid 60 fps; robotics motion reads fine at 24-30 fps and halves bitrate.
• Static title cards and captions compress to nearly nothing; noisy handheld

footage does not — use a tripod, lock exposure.

• Verify the final file on the actual spec: duration ≤ 180 s, size ≤ 20 MB,

extension among the allowed three, plays in a stock player.

Anonymization of footage

Video leaks identity faster than text (2026-cycle double-anonymous rules):

• Sweep frames for lab posters, door signs, safety-vest logos, license plates,

reflections in glass, and named equipment tags; blur or reshoot.

• No voice-over (voiceprints identify); no bystanders' faces without masking.
• Strip container metadata (ffmpeg -map_metadata -1 on the final pass) — the

encoder tag can carry a username.

Project pages and external links

A public project page is standard robotics practice after acceptance but a

liability during double-anonymous review:

• If the page already exists (earlier arXiv release), do not link it from the

submission; reviewers finding it on their own is tolerated, you leading them

there is a violation pattern.

• An anonymized mirror (no lab name in domain or footer) may be linked if the

cycle's rules permit external material at all — reread the current CFP.

• Plan the public page for acceptance week: paper, final video, artifact link,

BibTeX with the IEEE Xplore DOI (icra-camera-ready stages this flip).

Windows and logistics

In the 2026 cycle videos were accepted only during two windows (Aug 5-Sep 9 and

Sep 17-22, 2025, 23:59 Pacific) — note the second window closes after the paper

deadline, a grace period for editing but not for shooting. Bank raw footage

throughout the experimental campaign (icra-reproducibility ties clips to trial

logs); deadline-week shooting produces both worse videos and worse anonymity

discipline. Verify the current year's windows and specs on the year-site before

scheduling the edit.

Pre-upload checklist

1. Paper passes the self-containment test with video deleted.
2. Storyboard beats present: uncut core, labeled speeds, comparison, failure.
3. Trial conditions on screen match table conditions in the paper.
4. Spec check: ≤ 180 s, ≤ 20 MB, allowed container, plays everywhere.
5. Anonymity sweep of frames, captions, metadata complete.
6. Upload inside the posted window; confirmation archived.

Output format

[Self-containment] pass / fail — claims living only in video: <list>
[Storyboard audit] uncut-core / speed-labels / comparison / failure: y/n each
[Spec] <s> s, <MB> MB, <container> — compliant: y/n
[Anonymity] frame sweep, captions, metadata: clean / leaks: <list>
[Window] target window <dates> — footage ready by: <date>
[Cross-reference] video trials ↔ paper tables consistent: y/n