brycewang-stanford ICRA Review Process
ICRA review is run not by ad-hoc program committees but by a standing RAS Conference Editorial Board (CEB) structure operating inside PaperPlaza. The mechanics differ enough from both journals and ML conferences that authors mis-model them — most consequentially around the rebuttal that never comes.
The pipeline
Sep 15 PaperPlaza deadline
│
▼ Editor (area lead) assigns paper to an Associate Editor (AE)
│ by keyword match from your chosen PaperPlaza topics
▼
Oct-Dec AE recruits 2-3 reviewers; reviews written on the standard
│ RAS form (contribution, technical soundness, clarity,
│ experimental validation, video comments)
▼
Dec-Jan AE writes a summary report reconciling the reviews and
│ recommends; Editors + program committee calibrate across areas
▼
Jan 31 accept or reject — no conditional accepts, no shepherding,
│ and in the classic ICRA flow NO author rebuttal phase
▼
Mar camera-ready with voluntary reviewer-driven revisions
Two structural notes. First, keyword choice at submission is not paperwork — it routes the manuscript to an Editor's area and thereby to the reviewer pool most or least sympathetic to it. Second, reviewer identity practice: reviewers are anonymous to authors, and from the 2026 cycle authors are anonymous to reviewers too (double-anonymous; before 2025 inclusive, single-blind). Policies are set per cycle — confirm on the current year-site.
No rebuttal: consequences, not complaints
The classic ICRA pipeline gives authors no reply channel between submission and decision (individual cycles have experimented; verify the current one). Plan for it:
- The paper must anticipate objections in advance — the limitations
paragraph and pre-answered comparisons (
icra-writing-style) do the job a rebuttal does elsewhere. - Misunderstandings are terminal. If a reviewer misreads the method, no one corrects them before the AE summarizes. Ambiguity is therefore a first-order risk, worth spending page budget to eliminate.
- The video is your asynchronous rebuttal. It answers "is this real / is it
cherry-picked / how fast" before those objections are typed
(
icra-supplementary). - Reviews arrive with the decision, and for accepted papers they function as
camera-ready guidance (
icra-author-responsecovers using them well).
What the review form rewards
| Form dimension | What strong looks like at ICRA |
|---|---|
| Technical soundness | assumptions stated; math checkable; no hidden human-in-the-loop |
| Experimental validation | evidence rung matches claims; trials, dispersion, failure analysis |
| Novelty/contribution | mechanism-level delta over named prior systems |
| Clarity | figure chain self-sufficient; task-first narrative |
| Multimedia | uncut executions, labeled speeds, failure shown |
Borderline papers typically die on validation ("promising but insufficient experimental support") rather than novelty — allocate effort accordingly.
Where author leverage actually exists
With no rebuttal, influence concentrates before upload:
- Keyword selection — the only routing input you control; choose for the reviewer pool you want, within honesty.
- The video — the one channel that reaches reviewers with evidence after the PDF freezes their reading order but before they form judgments.
- Ambiguity elimination — every sentence a reviewer could misparse is a decision risk with no correction mechanism.
- The limitations paragraph — converts predicted objections from "flaw found" to "flaw acknowledged," a materially different review sentence.
Calibrating expectations
- ICRA is large (thousands of submissions; Vienna 2026 drew 8,000+ attendees); historical acceptance rates have floated in the 40-45% band, far gentler than ML flagship venues but with high variance across areas. Treat any specific rate as unverified until the year's numbers are published.
- Acceptance is to the program as a whole; presentation format (interactive/ poster session structure) is a program decision, not a quality tier ranking in the proceedings record.
- Decisions are binary. "Reject with encouraging reviews" is the common fate of good-but-thin papers; the correct read is "fund one more experiment rung and resubmit" (IROS in March, RA-L any time).
Contrast with RA-L review
Authors choosing between routes (see icra-topic-selection) should model the
review differences, not just the deadlines:
- RA-L is journal review: an AE plus reviewers, a first decision targeted within ~3 months, and crucially a revise-and-resubmit outcome that lets authors answer reviewers with a response letter — the interaction ICRA lacks.
- RA-L decisions bind publication in the journal; conference presentation is a separate transfer step afterward.
- The reviewer pools overlap heavily; a paper rejected at ICRA and sent to RA-L unchanged may meet the same reviewer with the same objection — revise first.
Reading a decision packet
- Sort comments into: factual error (yours), misunderstanding (their reading, your ambiguity), scope demand (wants a different paper), and taste.
- Reject case: fix the first two categories before any resubmission clock; a misunderstanding at ICRA predicts the same misunderstanding at IROS.
- Accept case: triage which comments to honor in camera-ready given the
8-page-total final limit (
icra-camera-ready). - Update your model: which claims drew fire predicts how to scope the next submission's evidence ladder.
Output format
[Stage] pre-submission / under review / decision received
[Pipeline model] AE area = <keyword>, predicted reviewer pool concern: <one line>
[No-rebuttal mitigation] top 3 objections pre-answered in draft: <list>
[Decision triage] factual / misunderstanding / scope / taste = <counts>
[Next move] camera-ready plan | resubmission target + what changes