rMuscle: Robotic Muscle Memory for Efficient Vision-Language-Action Model Inference
Kaijun Zhou, Zhiyang Li, Le Chen, Jinyu Gu
cs.RO · cs.AI
Quick verdict
Cross-task caching speeds VLA inference 1.20-1.42x at matched success rates; solid systems work, but limited to repetitive tasks and lacks statistical testing.
Summary
The paper presents rMuscle, a real-time inference framework for Vision-Language-Action (VLA) models deployed in repetitive industrial settings. Building on measurements showing that repeated executions of the same task yield similar observations, trajectories (path cosine ~0.994-0.995), FFN inputs/outputs (98.5%/81.0% similarity), and important-neuron sets (86% same-task overlap vs 71% across tasks), rMuscle introduces a dual-phase cache: a Context Cache that reuses cached FFN outputs for visual tokens whose inputs drift little (recomputing the top rho_p=0.4), and an Action Cache that loads weights only for predicted important neurons (rho_n=0.5) during memory-bound denoising, with dense anchor steps every ~5 steps. Cache management stores compact inputs, reconstructs expanded states online during robot execution (~25 ms/entry), retrieves references via intermediate ViT features plus joint-state history within a sliding window (T=4, L=4), and falls back to dense execution when visual similarity drops below 0.8. On pi0.5, GR00T N1.6, and X-VLA across 40 LIBERO and 50 RoboTwin tasks plus two physical tasks, rMuscle achieves 1.20-1.29x (RTX 4090) and 1.23-1.42x (Jetson Thor) end-to-end speedup over the FlashRT engine while matching success rates (35.2% RoboTwin average; 76%/84% physical), with under ~3 GB extra GPU memory. Main limitations: per-task offline profiling (200-800 inference calls, 8-32 reference episodes), applicability restricted to repetitive tasks, and success-rate comparisons reported without variance or significance testing.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Repeated executions of the same task share FFN inputs/outputs (98.5%/81.0% similarity) and important-neuron sets (86% same-task vs 71% cross-task overlap), making cross-execution computation and weight-access reuse viable.
- A dual-phase cache — Context Cache for compute-bound VLM prefill and Action Cache for memory-bound denoising — delivers 1.20-1.29x end-to-end speedup on RTX 4090 and 1.23-1.42x on Jetson Thor over the FlashRT inference engine.
- rMuscle matches vanilla policy success rates on LIBERO, RoboTwin, and two physical tasks (35.2% RoboTwin average; 76% ALOHA pick-and-place; 84% conveyor packing).
- Online cache reconstruction, sliding-window prefetching, and group-shared neuron masks keep overhead negligible: under ~3 GB extra GPU memory on a 24 GB card and cache management off the inference critical path.
- Relative speedup increases with model scale, reaching 1.6x (RTX 4090) and 1.9x (Thor) for the pi0.5-XXL configuration.
Strengths
Quantified motivation from workload characterization
Section 3 measures trajectory cosine similarity of 0.994-0.995, FFN input/output similarity of 98.5%/81.0%, and important-neuron overlap of 86% same-task vs 71% cross-task; Figure 5 shows selecting rho_p=0.4 of visual tokens captures ~80% of FFN output change, directly grounding the design constants rather than asserting them.
Bottleneck-matched dual design corroborated by ablation
Table 1 roofline separates compute-bound prefill (AI 891 FLOP/B) from memory-bound denoising (AI 34); the Figure 19 ablation confirms Context Cache alone accounts for 87% of RTX 4090 latency reduction while Action Cache alone speeds Thor denoising 1.49x, matching the hardware profiles.
Broad evaluation across models, tasks, and hardware
Careful cache-management engineering with measured overhead
Scaling trend and candid failure analysis
Weaknesses
No statistical treatment of success rates
Narrow applicability with unquantified profiling cost
Modest, unevenly distributed gains
Low-ceiling benchmark regime and thin physical evidence
No artifact release
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 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 rMuscle and what problem does it solve?
rMuscle is an inference-acceleration framework for Vision-Language-Action (VLA) robot policies. In structured settings like factories, robots repeat similar tasks, and the model's internal computations repeat too; rMuscle caches reusable FFN outputs and neuron-activation patterns from past executions, then reuses them at inference time to cut latency without retraining or quantizing the model.
How much speedup does rMuscle achieve?
Per the paper, rMuscle speeds up pi0.5, GR00T N1.6, and X-VLA by 1.20-1.29x end-to-end on an RTX 4090 and 1.23-1.42x on Jetson Thor versus the FlashRT inference engine — for example, pi0.5 drops from 46.0 ms to 35.5 ms per policy call on the 4090 (about 28 Hz) — while keeping task success rates unchanged in simulation and on physical robots.
Does rMuscle's caching reduce robot task success rates?
The authors report success rates equal to dense (uncached) execution: 35.2% average across 50 RoboTwin tasks, matched LIBERO averages across three models, and 76%/84% on two physical tasks (dual-arm bottle pick-and-place and conveyor furniture packing). Caveats: no statistical significance tests are reported, and two RoboTwin tasks lose 2-3 points when observation similarity sits near the 0.8 fallback threshold.
What are rMuscle's main limitations?
It requires per-task offline profiling (200-800 recorded inference calls and 8-32 reference episodes) and assumes repetitive work at a fixed workstation; for novel or highly variable tasks it falls back to dense execution and gains little. No code release is mentioned, and success-rate comparisons are reported without variance or significance analysis, so the reported gains apply specifically to repetitive-task deployments.
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.
- Figure-embedded data could not be verified numerically: LIBERO per-suite success rates (Figure 17) and several RoboTwin rows (Table 4 shows only 9 of 50 tasks) appear only as figure content.
- Figures 4, 6, 9, 11-14 are images; the qualitative evidence (patch localization, temporal-locality plot, physical setups) is assessed from captions and in-text numbers only.
- The supplied abstract excerpt is truncated mid-sentence ('maintainin...'); claims were checked against the full embedded text instead.
- arXiv v1 only; no artifact/appendix statement beyond the main text was available.
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.