Writing study: RoboMME: Benchmarking and Understanding Memory for Robotic Generalist Policies

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Yinpei Dai, Hongze Fu, Jayjun Lee, Yuejiang Liu, Haoran Zhang, Jianing Yang, Chelsea Finn, Nima Fazeli, Joyce Chai · Official ICML 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Datasets, benchmarks & measurement (primary).

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF

55 physical pages

Before references / appendix

≈ 9.21 page-equivalents

Main figures / tables

4 / 4

Attached appendix figures / tables

23 / 28

Attached appendix span

≈ 41.91 page-equivalents

Before references / appendix: ≈ 16.75%References / intervening material: ≈ 7.05%Attached appendix: ≈ 76.2%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level heading	Starts on PDF page	Estimated span	Share of entire PDF
1. Introduction	1	≈ 1.44 pp.	2.62%
2. Related Work	2	≈ 0.6 pp.	1.09%
3. RoboMME Benchmark	3	≈ 1.74 pp.	3.16%
4. Memory-Augmented Manipulation Policies	5	≈ 1.67 pp.	3.04%
5. Experiments	6	≈ 2.97 pp.	5.4%
6. Conclusion and Future Work	9	≈ 0.26 pp.	0.47%
Acknowledgments	10	≈ 0.08 pp.	0.15%
Impact Statement	10	≈ 0.09 pp.	0.16%
References	10	≈ 3.88 pp.	7.05%
A. Model Architectures in the MME-VLA Suite	14	≈ 5 pp.	9.09%
B. Experiment Implementation Details on RoboMME	19	≈ 11 pp.	20%
C. Full Results on RoboMME	30	≈ 7.25 pp.	13.18%
D. Real Robot Experiment	37	≈ 2 pp.	3.64%
E. Task Description	39	≈ 16.66 pp.	30.29%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Table 1 · p. 4
• Table 2 · p. 5
• Figure 2 · p. 6
• Table 3 · p. 8
• Figure 3 · p. 9
• Figure 4 · p. 9
• Table 4 · p. 9

Attached appendix

• Table 5 · p. 19
• Figure 5 · p. 20
• Figure 6 · p. 21
• Figure 7 · p. 22
• Table 6 · p. 24
• Table 7 · p. 25
• Table 8 · p. 26
• Table 9 · p. 27
• Table 10 · p. 28
• Figure 8 · p. 29
• Table 11 · p. 30
• Table 12 · p. 31
• Table 13 · p. 32
• Table 14 · p. 32
• Table 15 · p. 34
• Table 16 · p. 35
• Table 17 · p. 35
• Table 18 · p. 36
• Figure 9 · p. 37
• Figure 10 · p. 38
• Figure 11 · p. 39
• Figure 12 · p. 40
• Table 19 · p. 40
• Figure 13 · p. 41
• Table 20 · p. 41
• Figure 14 · p. 42
• Table 21 · p. 42
• Figure 15 · p. 43
• Table 22 · p. 43
• Figure 16 · p. 44
• Table 23 · p. 44
• Figure 17 · p. 45
• Table 24 · p. 45
• Figure 18 · p. 46
• Table 25 · p. 46
• Figure 19 · p. 47
• Table 26 · p. 47
• Figure 20 · p. 48
• Table 27 · p. 48
• Figure 21 · p. 49
• Table 28 · p. 49
• Figure 22 · p. 50
• Table 29 · p. 50
• Figure 23 · p. 51
• Table 30 · p. 51
• Figure 24 · p. 52
• Figure 25 · p. 53
• Figure 26 · p. 54
• Table 31 · p. 54
• Figure 27 · p. 55
• Table 32 · p. 55

Argument and organization

1. Problem and contribution

   Distinguishes tasks that merely contain temporal sequences from tasks that actually require unavailable past information, motivating four different memory demands.

2. Argument sequence

   Task design and construction precede a representation-by-integration taxonomy; evaluation questions then separate symbolic sufficiency, human performance and task-dependent design effects.

3. Evidence and verification

   The human reference uses incremental VideoQA with an oracle low-level planner, whereas robot policies execute control; the text explicitly identifies that protocol distinction.

Problem, claim and contribution

Distinguishes tasks that merely contain temporal sequences from tasks that actually require unavailable past information, motivating four different memory demands.

Introduction/related work and Figure 1, pp. 1-2 · Main paper

Argument and section order

Task design and construction precede a representation-by-integration taxonomy; evaluation questions then separate symbolic sufficiency, human performance and task-dependent design effects.

Sections 3-5, pp. 3-9 · Main paper

Figures, captions and visual explanation

Time-indexed examples pair four memory questions with concrete tasks; the full task matrix separates oracle subgoals, generated subgoals and neural memory variants rather than mixing their averages.

Figure 1, p. 2; Table 3, p. 8; both PNGs inspected · Main paper

Tables, comparisons and evidence

The human reference uses incremental VideoQA with an oracle low-level planner, whereas robot policies execute control; the text explicitly identifies that protocol distinction.

Section 5, Q3, p. 8 · Main paper

Appendix and reproducibility

Task sheets state difficulty, information phases and early-failure rules; token-budget and attention analyses disclose a smaller 160-episode diagnostic sample.

Appendix C.5, p. 35; E.16, p. 55 · Main paper

Clarity and writing conventions

Negative and task-dependent findings are part of the main story: no design wins every task and oracle symbolic guidance still leaves manipulation bottlenecks.

Section 5, Q2-Q4, p. 8 · Main paper

What to reuse—and what not to infer

Transferable lesson: Make the benchmark's construct unavoidable by design, then cross representation and integration choices while separating oracle and executable protocols.

Boundary: The human score is not a matched embodied-control ceiling; attention allocation and modulation statistics suggest usage but do not by themselves prove causal memory dependence.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-focused distributed manuscript scan, not exhaustive technical/proof validation. Sections 1-5 and discussion, pp. 1-9: history-dependent task design, benchmark construction, memory representation/integration taxonomy and question-led evaluation.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendices A-E, pp. 14-55: model variants, training/evaluation details, full results/token analyses, robot tests and sixteen task-specific protocols.

Visually inspected PDF pages

2, 8

This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.

• Published ICML 2026 proceedings PDF, PMLR volume 306; retrieved 2026-10-07. Public OpenReview decision notes were not independently reverified. · 55 pages