---
name: brycewang-stanford/corl-supplementary
source: https://app.decimal.ai/s/brycewang-stanford-corl-supplementary@1/SKILL.md
source_sha256: fe19d710935e
---

# CoRL Supplementary Material

CoRL supplementary material splits into three channels with different rules and
different reader guarantees: the **appendix** (same PDF, optional reading), the
**supplementary upload** (files and the venue's signature overview video), and —
after acceptance — **external hosting**, because the PMLR proceedings pipeline
does not take videos at all. Content placed in the wrong channel either never
gets seen or violates a rule.

## Channel rules (2026 cycle, corl.org, read 2026-07-08)

| Channel | 2026 rule | Reader guarantee |
|---|---|---|
| Appendix | Optional; follows references in the same PDF; uncounted against the 8 pages | Reviewers are **not required** to read it |
| Supplementary file(s) | Encouraged for further details; video ≤ 250 MB (strict), ~3 min suggested | Also optional reading; must be anonymous |
| External links during review | Only if fully anonymous — the URL included | Assume most reviewers won't click |
| Camera-ready / PMLR | **No video as supplementary**; host video/code/data externally, link from the main text or a project page | Permanent record; plan durable hosting |

The uniform consequence of "optional reading": nothing decision-critical may live
outside the 8-page body. The body must stand alone; supplementary material
corroborates, it never carries.

## The video: evidence, not advertisement

For a learned-policy paper the video has a specific evidentiary job — showing
that printed success rates correspond to real behavior. Reviewers at this venue
watch with a checklist mindset:

- **Uncut executions.** At least a few complete episodes per headline task with
  no cuts inside an attempt; jump-cut montages are read as concealment.
- **Playback honesty.** Label every speed change on screen ("4x"); real-time
  segments for anything you call reactive or real-time.
- **Failure cases.** Include representative failures matching your Limitations
  section — the pairing is checked, and showing them buys credibility that
  no success reel can.
- **Regime labels.** Mark sim vs real footage explicitly; a photoreal simulator
  clip mistaken for hardware, then discovered, poisons the whole packet.
- **Anonymity.** No faces, badges, lab signage, institution-branded robot
  livery, or watermarks; mute or scrub audio (voices identify people).

```bash
# Fit the strict 250 MB cap without shredding quality (H.264 target ~8 Mbps)
ffmpeg -i raw_cut.mp4 -c:v libx264 -preset slow -b:v 8M -maxrate 10M \
       -bufsize 16M -vf scale=1920:-2 -an overview_anon.mp4   # -an strips audio
stat -c '%s bytes' overview_anon.mp4          # verify < 250*1024*1024
# Frame-level anonymity spot check: export stills every 5 s and eyeball them
ffmpeg -i overview_anon.mp4 -vf fps=1/5 frames/check_%03d.png
```

Suggested three-minute arc: task family and setup (20 s) → method in one
diagram (20 s) → uncut successes across tasks/objects (90 s) → failures +
limitations (30 s) → generalization or sim-to-real side-by-side (20 s).

## The appendix: overflow with a contract

Because reviewers may skip it, apply a simple contract test to every block you
move there: *if a reviewer never reads this, does any claim in the body lose its
support?* If yes, it stays in the body (or the claim softens).

Belongs in the appendix: full hyperparameter and architecture tables, per-task
result breakdowns behind body-level aggregates, extra qualitative rollouts,
prompt lists for language-conditioned systems, rig photographs, dataset
datasheets, extended related work.

Does not belong: the only statement of the evaluation protocol, the only
sim-to-real numbers, the definition of the success criterion, or the honest
version of a claim the body states more boldly.

## Code and data attachments during review

- Ship a self-contained, anonymized bundle: training config, evaluation scripts,
  environment definitions, and a small demo-data sample if the full set is heavy.
- Scrub before zipping — usernames in paths, git history, W&B project names,
  internal hostnames, license files naming the lab (`corl-submission` has the
  mechanical sweep).
- A `README` at the bundle root mapping files to paper tables costs ten minutes
  and is the difference between "code provided" and "code checked."

## Camera-ready transition (the PMLR pivot)

On acceptance the channels reorganize (deadline: October 12, 2026, 23:59 AoE for
the 2026 cycle):

1. Appendix merges into the end of the camera-ready PDF (`corl-camera-ready`).
2. The review-time video cannot follow the paper into PMLR — publish it on a
   project page or stable video host, de-anonymized, and cite the link in the
   main text.
3. Code/data move to their permanent public homes; prefer an archival DOI for
   datasets you expect others to benchmark against (`corl-artifact-evaluation`).

## Assembly checklist

```text
[ ] Body stands alone with supplementary deleted (contract test passed)
[ ] Video: uncut episodes, labeled speeds, failures included, regimes marked
[ ] Video ≤ 250 MB, audio stripped, frame-level anonymity check done
[ ] Appendix in the same PDF, after references, no load-bearing content
[ ] Code bundle anonymized, README maps files → tables
[ ] External links (if any) anonymous including the URL itself
[ ] Post-acceptance hosting plan drafted (PMLR takes no video)
```

Caps, encouragement language, and the PMLR handoff are all cycle-specific:
re-open https://www.corl.org/contributions/instruction-for-authors for the live
year before finalizing the bundle.