vla.simd: Efficient CPU Inference for Language-Conditioned Manipulation
Khanh D. Nguyen, Hoang M. Truong, An T. Le
cs.RO · cs.AI · eess.SY
Quick verdict
CPU SIMD engine with ~1.4x median speedup and IMPACT policy sustains 30 actions/s on Pi 5; rigorous timing analysis; sim success trails top baselines.
Summary
This paper presents vla.simd, a CPU inference engine for language-conditioned manipulation policies, together with IMPACT, an ACT-based policy that adds language conditioning at roughly 2.6% of per-query compute. The engine runs a packed-panel SIMD micro-kernel shared by linear layers, convolutions, and attention projections, selects register tiles offline per target CPU, and supports optional W8A8 int8 quantization. An action-supply analysis derives feasibility conditions for lagged execution (C1: d ≤ H) and time-aligned execution (C2: 2d ≤ H) at a 30 Hz controller, separating action-supply rate from feedback frequency. Across six policies and four CPUs (Apple M4, Intel i9-14900HX, AMD Ryzen 5 5500, Raspberry Pi 5), vla.simd achieves about 1.4× median speedup over compiled PyTorch (range 0.83–3.37×) with maximum action deviations of 8.8×10⁻⁵ robot units. After a 90 s thermal soak on the Pi 5, IMPACT supplies 33.5 actions/s in fp32 and 81.2 with int8 — the only language-conditioned policy in the evaluated set above 30 actions/s. GPU evaluation yields 76.4% mean LIBERO success, below published TurboVLA and SmolVLA results obtained under different protocols; instruction shuffling shows 91% selection of newly instructed goals. Limitations include single-seed simulation, 20-trial robot samples, unmeasured tail latency, and no comparison with production runtimes such as ONNX Runtime.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 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
- vla.simd achieves approximately 1.4× median speedup over compiled PyTorch across six policies and four CPUs while preserving fp32 numerical fidelity (max action deviation 8.8×10⁻⁵ robot units; Intel/AMD bitwise identical).
- IMPACT is the only language-conditioned policy in the evaluated set supplying at least 30 actions/s on the Raspberry Pi 5, delivering 33.5 actions/s in fp32 and 81.2 with int8 after a 90 s thermal soak.
- Action availability under sequential chunked execution is governed by feasibility conditions C1 (d ≤ H, lagged execution) and C2 (2d ≤ H, time-aligned execution) at a 30 Hz controller, separating action supply from feedback frequency.
- IMPACT reaches 76.4% mean LIBERO success without robot pretraining and selects newly instructed goals in 91% of 600 instruction-shuffling rollouts, with language conditioning adding only ~2.6% of per-query compute.
Strengths
Analytically useful timing framework
Section III derives separate feasibility conditions for lagged (C1: d ≤ H) and time-aligned execution (C2: 2d ≤ H) and explicitly distinguishes action-supply rate (feff = H/Tq) from feedback frequency — a correction directly applicable to any chunked-policy deployment, visualized for all six policies in Fig. 1.
Measurement hygiene unusual for reimplementation papers
Median-of-15 fresh-process timings, A–B–B–A interleaving to limit thermal drift (Section VI-A), fidelity gates (10⁻² absolute, 10⁻⁴ relative) verified before any timing is trusted, and bitwise-identical fp32 actions between Intel and AMD backends (Section VI-C).
Transparent handling of unfavorable results
Per-target tuning shown to matter, not assumed
Grounding on two real embodiments with CIs
Weaknesses
No production-runtime baseline
Feasibility budgets exclude the real latency path
Simulation results trail modern baselines under differing protocols
int8 task-performance validation is thin
Real-robot language attribution is confounded
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets GPU-free deployment of language-conditioned manipulation policies through three coupled contributions. First, **vla.simd**, a from-scratch CPU inference engine built on a packed-panel SIMD micro-kernel shared by linear layers, convolutions (via panel-wise im2col), and attention projections, with per-target register tiles (6×16 AVX2, 4×16 NEON) selected offline by Algorithm 1 under register-pressure constraints (Eqs. 2–4) and a SPILLS compile check. Second, **IMPACT**, a 60M-parameter ACT derivative (plus a 35.3M frozen T5-small encoder) that appends cached projected text tokens to the encoder and mean-pools them into per-stage FiLM coefficients, adding ~2.6% of per-query FLOPs (~88 GFLOP/query in the SO-101 configuration). Third, an **action-supply budget analysis** (Section III) giving feasibility conditions C1 (d ≤ H, lagged execution) and C2 (2d ≤ H, time-aligned execution with stale-action discard) at a 30 Hz controller. Headline claims: ~1.4× median speedup over compiled PyTorch across 28 policy–CPU pairs (range 0.83–3.37×) with fp32 fidelity (max deviation 8.8×10⁻⁵ robot units); IMPACT is the only language-conditioned policy in their set sustaining ≥30 actions/s on the Pi 5 (33.5 fp32 / 81.2 int8 after a 90 s thermal soak); 76.4% mean LIBERO success without robot pretraining; 91% instructed-goal selection under instruction shuffling; CPU deployment on SO-101 (IMPACT) and UR10e (SmolVLA).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- 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 vla.simd?
vla.simd is a CPU inference engine for vision-language-action (VLA) robot manipulation policies. It uses hand-optimized SIMD kernels shared across linear layers, convolutions, and attention projections, with per-CPU register tile selection and optional int8 quantization. The paper reports about 1.4× median speedup over compiled PyTorch across six policies and four CPUs while matching fp32 reference outputs to within 8.8×10⁻⁵ robot units.
What is the IMPACT policy?
IMPACT (Instruction-Modulated Perception and Action Chunking with Transformers) extends the ACT manipulation policy with language conditioning: a frozen T5-small text encoder runs once per instruction, its cached tokens are appended to the transformer encoder, and a mean-pooled embedding produces FiLM scale/shift coefficients for each ResNet-18 stage. Language costs about 2.6% of per-query compute, and the policy predicts 50-action chunks from two camera views plus robot state.
Can vision-language-action policies run in real time on a Raspberry Pi 5?
According to this paper, most evaluated policies cannot, but IMPACT can: after a 90-second thermal soak it supplies 33.5 actions per second in fp32 and 81.2 with int8 on the Pi 5, above the 30 actions/s a 30 Hz controller consumes. It is the only language-conditioned policy in the paper's set meeting that budget; ACT, which has no language input, also passes.
How successful is IMPACT on standard benchmarks?
On LIBERO simulation, IMPACT reaches 76.4% mean success across four suites without robot pretraining (78.6% with checkpoint selection). This is below published TurboVLA (97.7%) and SmolVLA (87.3%) results, though those use different observation inputs and execution protocols, so they serve as context rather than a controlled comparison. In instruction-shuffling tests it picked the newly instructed goal in 91% of 600 rollouts.
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 are only partially recoverable from embedded caption text; marker placements and axis readings could not be verified visually.
- Several key comparators (TurboVLA, vla.cpp, DEM, Lite VLA) are recent arXiv preprints; their published numbers were taken at face value.
- No network access: the project page (vla-simd.github.io), including any code release, could not be checked; codeAvailable reflects the paper text only.
- Extraction appears complete (abstract through references), but any supplementary material beyond the PDF body was unavailable.
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.