StateMem: Single-State Residual Memory with Adaptive Inference for Vision-Language-Action Policies
Wenzhuo Li, Qiongfeng Shi, Yi Zhou
cs.RO · cs.LG
Quick verdict
Single-token residual memory plus error-gated prefix caching for VLAs; real memory-task gains, but margins lack error bars and efficiency savings are modest.
Summary
StateMem augments a π0.5-based vision-language-action policy with a persistent single memory token that is updated at every step by a low-rank projection of the prediction error between a lightweight predictor's estimate of the current perceptual feature and its actual value. The same prediction error, computed against the cached anchor, routes inference: consistent observations reuse the VLM prefix KV cache, while a distilled fast-correction adapter compensates for stale cache and a training-free EMA controller adapts the routing threshold via periodic action audits. On LIBERO the method reaches 97.6% average success (best among compared methods, +0.7 over π0.5) while cutting prefix refreshes to 79.75%, improving latency (150.78 vs 156.36 ms) and throughput (6.64 vs 6.40 Hz). On RoboMemArena it reports the best results among single-VLA baselines (average TSR 26.8%, CSR 44.4%), though hierarchical VLM+VLA systems (PrediMem, MemER) remain ahead, and refresh stays at 98.9% there, making routing latency-neutral-to-negative. Across six real-world xArm7 tasks it averages 65% success, +21 points over π0.5, driven almost entirely by the two memory-dependent tasks where π0.5 scores zero. Main limitations: no error bars, seeds, or significance tests anywhere; modest absolute efficiency gains; single-robot, 20-trial real-world protocol; and hyperparameters deferred to code.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single memory token updated by low-rank residual prediction error improves memory-dependent manipulation, achieving the best RoboMemArena results among single-VLA baselines (avg TSR 26.8%, CSR 44.4%).
- StateMem reaches 97.6% average success on LIBERO, the highest among compared methods and +0.7 over π0.5.
- Prediction-error routing with a training-free online controller cuts LIBERO prefix refreshes to 79.75% while slightly improving latency (150.78 ms) and throughput (6.64 Hz) without success-rate loss.
- Across six real-world tasks StateMem averages 65% success, +21 points over π0.5, with gains concentrated in the two memory-dependent tasks where π0.5 scores zero.
Strengths
Component-isolating ablations support the mechanism
Table IV: removing residual memory drops TSR 27.7→22.8 and CSR 45.4→39.0; removing prediction error drops them to 25.0/42.9 — the error-driven residual update, not merely added parameters, drives the memory gains.
Unusually candid negative results on its own efficiency mechanism
Sec. IV-D reports that on RoboMemArena routing is net-negative (98.9% refresh; 137.9 vs 135.1 ms latency) and the no-routing variant matches success (26.8/44.5); the limitation is restated in Sec. V rather than hidden.
Controlled history-retention probe
One pipeline across three settings with documented costs
Weaknesses
Zero statistical reporting
Efficiency benefit is small and regime-dependent
Real-world protocol weak for headline claims
Trails hierarchical memory systems and regresses in two categories
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
StateMem (arXiv 2609.22684, cs.RO) is a method paper that augments a π0.5-based vision-language-action policy (PaliGemma-2B VLM + Gemma-300M action expert) with three coupled mechanisms: (1) a single persistent memory token updated each step by a low-rank residual (A ∈ R^{r×dp}, B ∈ R^{dm×r}) of the prediction error between a lightweight predictor's estimate of the current perceptual feature and its measured value (Sec. III-B, Eq. 4–9); (2) adaptive VLM prefix routing that reuses the cached KV prefix when the anchor-conditioned prediction error is low, with a distillation-trained fast-correction adapter compensating for stale caches (Sec. III-C, Eq. 10–17); and (3) a training-free EMA threshold controller with periodic cached-vs-full-refresh action audits (Sec. III-D, Eq. 18–22). Headline claims: 97.6% average success on LIBERO (+0.7 over π0.5) at a 79.75% prefix refresh rate; best single-VLA results on RoboMemArena (avg TSR 26.8%, CSR 44.4%); and 65% average success across six real-world tasks, +21 points over π0.5.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is StateMem and how does it differ from MemoryVLA?
StateMem is an add-on for vision-language-action robot policies that keeps task history in a single memory token updated by prediction error through low-rank residuals, instead of MemoryVLA's external memory bank with explicit storage, retrieval and consolidation. On the RoboMemArena benchmark it reports average TSR 26.8%/CSR 44.4% versus MemoryVLA's 15.0%/35.3%, and trains faster (9.37 vs 15.22 seconds per iteration).
How much compute does StateMem's adaptive prefix routing actually save?
On LIBERO it lowers the VLM prefix refresh rate from 100% to 79.75%, cutting per-query latency from 156.36 to 150.78 ms and raising throughput from 6.40 to 6.64 Hz. On RoboMemArena the refresh rate stays at 98.9% and latency is slightly worse than always refreshing (137.9 vs 135.1 ms), so the benefit depends on how often observations stay consistent.
What results does StateMem report on the LIBERO benchmark?
97.6% average success across the four suites — Spatial 98.6, Object 99.2, Goal 98.2, Long 94.4 — the highest average among the compared methods, +0.7 points over π0.5 (96.9%), evaluated with 50 trials per task.
Is code available for StateMem?
Yes, the paper links anonymized source code (anonymous.4open.science/r/Anonymous0311) and documents the recipe: initialized from π0.5, two-stage training of 40K memory steps plus 20K fast-correction steps on four RTX A6000 GPUs in about five days. Most hyperparameter values are only in that code.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–5 appear only as captions and garbled axis text; per-task figure content could not be analyzed, though all cited numbers come from the cleanly extracted tables and body text.
- Hyperparameters (rank r, loss weights, controller constants) are deferred to the anonymized code and could not be verified from the text.
- Table I baseline provenance (literature-quoted vs re-run under the authors' protocol) cannot be verified against the original papers from this text alone.
- The text appears complete (all sections and tables present), but any appendix, if one exists, was not included in the extraction.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.