thedixitjain CoRL Related Work
Robot learning is a confluence field: any given CoRL paper inherits from at least three literatures moving at different speeds. Reviewers here are typically expert in one lane and conversant in the rest — so a related-work section that covers only the authors' home lane reads as blinkered to a third of the pool.
The three lanes, plus two overlays
| Lane | What reviewers check | Typical miss |
|---|---|---|
| Robot-learning lineage (CoRL, RSS, ICRA/IROS learning papers) | Do you know the last 2–3 years of the task family you claim? | Citing only your lab's chain of prior work |
| ML methods (NeurIPS/ICLR/ICML) | Is the algorithmic idea actually new, or a known method re-embodied? | Renaming an established technique with a robotics word |
| Classical robotics (planning, control, estimation) | Does a non-learning baseline solve your task? Why learn at all? | Dismissing model-based methods without citation or comparison |
| Overlay: foundation-model / VLA wave | Position relative to the current generation of pretrained robot policies | Comparing against a wave that is two generations stale |
| Overlay: concurrent arXiv | Preprint culture is aggressive here; overlap appears mid-review | Silence on a widely known concurrent preprint |
Practical breadth check: your citation list should contain entries from at least two recent CoRL volumes, at least one ML venue, and at least one classical robotics source — or an explicit reason why a lane is genuinely empty.
Handling velocity: the field moves during your review
Between the late-May CoRL deadline and September decisions, the arXiv state of your subfield will change. Norms to apply:
- Work that appeared publicly after your submission cannot be held against your novelty; if a reviewer raises it, note the dates factually in the rebuttal and offer a comparison in revision if it is genuinely relevant.
- Work that was on arXiv well before the deadline is fair game even without a venue stamp — this community treats known preprints as citable context. Sweep arXiv thoroughly in deadline month; "it wasn't published yet" does not excuse missing a preprint your reviewers all saw on social media.
- For close concurrent work, add a labeled paragraph ("Concurrent work") that differentiates on approach and evidence rather than priority claims.
Positioning moves that work at this venue
- Differentiate on evidence regime, not just method. "Prior work demonstrates X in simulation; we provide the first real-robot evaluation with a measured transfer gap" is a legitimate and well-received axis here.
- Name the assumption you remove. "Existing approaches assume known object poses; we operate from raw RGB" — a robotics-meaningful delta beats an architecture tweak in this pool's eyes.
- Give the task-family history one honest sentence of state-of-the-art, including the best non-learning result. Erasing the classical solution to make the learning look necessary is a detectable and damaging move.
- Baselines follow citations: any method you describe as closest related
work should appear in your tables, or the text should say why it cannot
(
corl-experimentscovers fairness mechanics).
PMLR citation hygiene — including the year-offset trap
CoRL proceedings publish through PMLR, and two mechanical errors recur:
- The year offset. PMLR volumes sometimes carry a publication year after
the conference year — e.g., the CoRL 2021 best paper (Chen et al.,
in-hand re-orientation) is
PMLR v164, 2022, and SayCan from CoRL 2022 appears with 2023 pagination metadata. Cite the conference edition in prose ("at CoRL 2021") even when the BibTeX year says otherwise, and keep the pair consistent across your references. - Venue misattribution. Neighboring PMLR volumes belong to different conferences; copying a BibTeX entry from a random site routinely mislabels a CoRL paper as appearing at another venue or vice versa. Resolve every robot learning citation against the volume index (v205 = CoRL 2022, v229 = 2023, v270 = 2024, v305 = 2025).
@inproceedings{chen2022inhand,
title = {A System for General In-Hand Object Re-Orientation},
author = {Chen, Tao and Xu, Jie and Agrawal, Pulkit},
booktitle = {Proceedings of the 5th Conference on Robot Learning (CoRL 2021)},
series = {Proceedings of Machine Learning Research},
volume = {164},
pages = {297--307},
year = {2022}, % PMLR publication year != conference year — keep both visible
publisher = {PMLR}
}
Prefer arXiv-to-PMLR upgrade passes before camera-ready: many robot-learning papers you cited as preprints during writing will have acquired proceedings entries by October.
Double-anonymous self-citation
- Cite your own prior work in the third person, exactly as you would a stranger's; the giveaway is not the citation but the possessive framing ("building on our platform [12]").
- If the paper visibly extends a system with a recognizable name, do not amputate the citation — anonymity through omission breaks the scholarly record; third-person distance is the sanctioned tool.
- Watch transitive leaks: citing an unreleased internal report, a thesis, or a dataset only your lab could hold identifies you as surely as a name.
Section-audit checklist
[ ] All three lanes represented (or absence argued), plus VLA-wave currency
[ ] ≥2 recent CoRL volumes cited within the claimed task family
[ ] Best classical/non-learning approach acknowledged, compared or excused
[ ] Concurrent-work paragraph drafted if a known preprint is close
[ ] Every "closest work" citation appears in the experiments, or is excused
[ ] PMLR volume/year pairs verified against volume indexes; no misattribution
[ ] Self-citations third-person; no possessive framing; no transitive leaks
Verify volume anchors and any new proceedings arrangements at https://proceedings.mlr.press/ and the live corl.org pages; this file's anchors were checked 2026-07-08.
Source: brycewang-stanford/Awesome-Journal-Skills → CoRL-Skills/skills/corl-related-work/SKILL.md