An T. Le

—

Unclaimed · AI-structured profileOpenAlex

Robot-learning researcher making vision-language-action models run faster via low-bit quantization and CPU inference

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-13) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

Not yet scored: this profile holds only two catalog papers and no verified publication-count or productivity metrics. This axis would measure total research output, venue spread, and publication volume over time once a fuller bibliographic record is linked.

Impact—

Not yet scored: citation counts, h-index, and downstream adoption are unavailable for this record. This axis would measure how far the two VLA-efficiency papers are cited, reused, or deployed by other robotics and systems groups.

Focus0/100

Scores 0 on focus: the field distribution splits exactly 50/50 between two labeled subfields — VLA model quantization under efficient inference, and VLA models for robotic manipulation. In practice the two halves are adjacent rather than divergent: both catalog papers target the same vision-language-action stack, one compressing it (W4A4, INT8) and one accelerating it on CPU. The score flags breadth across indexing categories, not a scattered agenda.

Collaboration44/100

Scores 44 on collaboration with 8 distinct coauthors in the record. The frequent-coauthor list — including Khanh D. Nguyen, Hung T. Ho, Quang D. Nguyen, Thanh Q. Duong, Ngan Le, Meng Guo, Vien A. Ngo, and Hoang M. Truong — suggests a stable, moderately sized core team rather than a wide external network. On both catalog papers Le sits in the last author position without the corresponding-author flag, a pattern consistent with senior or supervisory placement within that group.

Activity—

Not yet scored: no multi-year publication series is linked to this profile. This axis would measure publishing cadence and recency — for instance whether the two 2026 arXiv papers mark a sustained line of work or a new research direction.

Reproducibility55/100

Scores 55 on reproducibility, computed from a mean of 55 across 2 scored papers, both rated medium for repro-readiness. Evolve's reviews are consistent with that middle band: vla.simd is backed by rigorous timing analysis and concrete throughput figures (30 actions/s on a Raspberry Pi 5), and FoldQuantVLA reports honest statistics for its W4A4 pipeline, though its headline robot gain rests on small confounded trials.

Research summary

An T. Le works on robot learning with a distinct emphasis on making vision-language-action (VLA) models practical to deploy. The Evolve catalog holds two 2026 arXiv papers, in both of which Le is the last-listed author (not corresponding author). vla.simd builds a CPU SIMD inference engine with INT8 quantization and action chunking for language-conditioned manipulation; Evolve's review notes a ~1.4x median speedup and an IMPACT policy sustaining 30 actions/s on a Raspberry Pi 5, with rigorous timing analysis, though simulated success trails the top baselines. FoldQuantVLA targets native low-bit post-training quantization to W4A4 via consistent folding, Hadamard transforms, and GPTQ, delivered as native TensorRT INT4 plugins; the verdict credits credible speedups and honest statistics while flagging that the headline robot gain rests on small confounded trials. Bibliometrically, Le's record splits evenly (0.5/0.5) between VLA quantization and VLA manipulation subfields, with 8 distinct coauthors and a reproducibility score of 55 computed across those 2 scored papers; collaboration scores 44 and focus 0, reflecting a compact team and a two-subfield spread within one coherent theme. No institutional affiliation, country, or ORCID is attached to this profile, and the output, impact, and activity axes currently lack sufficient data to score.

Collaboration pattern

The record shows a compact collaboration pattern: 8 distinct coauthors, all recurring. The frequent-coauthor list (Khanh D. Nguyen, Hung T. Ho, Quang D. Nguyen, Thanh Q. Duong, Ngan Le, Meng Guo, Vien A. Ngo, Hoang M. Truong) reads as one stable core team behind both 2026 papers, where Le takes the last author position without the corresponding-author flag. No cross-institution or international collaboration data is available.

Research keywords

vision-language-action modelspost-training quantizationW4A4 quantizationHadamard transformGPTQINT8 quantizationCPU inferenceSIMD micro-kernelsaction chunkingrobotic manipulationTensorRT pluginsedge deployment

Highlights

FoldQuantVLA: native W4A4 post-training quantization for VLA models
Uses consistent folding with Hadamard transforms and GPTQ to push vision-language-action models to W4A4, shipped as native TensorRT INT4 plugins. Evolve's review credits credible speedups and honest statistics, but flags that the headline robot gain rests on small, confounded trials.
vla.simd: CPU-only VLA inference sustaining 30 actions/s on a Raspberry Pi 5
An INT8-quantized CPU SIMD engine with SIMD micro-kernels and action chunking for language-conditioned manipulation, delivering a ~1.4x median speedup; the IMPACT policy sustains 30 actions/s on a Pi 5. Evolve notes rigorous timing analysis, while simulated success trails the top baselines.
A coherent compress-and-deploy agenda for VLA models
Both 2026 papers come from one agenda: FoldQuantVLA compresses VLA models (W4A4, TensorRT INT4) while vla.simd accelerates them on commodity CPUs (INT8, ~1.4x median speedup), with Le in the last author position on each and the profile's 8 distinct coauthors clustered around this line of work.

Timeline

2026Publishes vla.simd, a CPU SIMD inference engine for language-conditioned manipulation, as last author on arXiv.
2026Publishes FoldQuantVLA, native W4A4 quantization of vision-language-action models via consistent folding, as last author.
2026Both catalog papers appear in the same year, forming a compress-and-deploy agenda around VLA efficiency with a core team of 8 distinct coauthors.

Papers in the Evolve catalog

Frequent co-authors

Khanh D. Nguyen×2 Hung T. Ho×1 Quang D. Nguyen×1 Thanh Q. Duong×1 Ngan Le×1 Meng Guo×1 Vien A. Ngo×1 Hoang M. Truong×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

Bibliometric scores are computed from OpenAlex-derived bibliometric data only; no citation, h-index, or venue-tier data is included for this record.
No institutional affiliation, country, or ORCID is attached; the profile is name-resolved from public bibliographic data, so identity cannot be fully confirmed.
Output, impact, and activity dimensions are unscored because the underlying data is insufficient; the reviewed catalog contains only two arXiv papers from 2026.
Both papers carry medium repro-readiness; quantitative claims are taken from Evolve's reviews and have not been independently verified.