Kunyu Peng
—
Robotics and computer vision researcher: VLA manipulation, LiDAR scene understanding, and reliability benchmarking
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-14) 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: Evolve has no publication-count or venue-volume metrics for this researcher yet. This axis would measure total research output, such as the number of papers per year and the types of venues, across the full bibliographic record.
Impact—
Not yet scored: citation counts, h-index, and other influence indicators are unavailable in the current data. This axis would measure how widely this researcher's papers are cited and taken up by the robotics and computer vision communities.
Focus0/100
Scores 0, reflecting topical breadth rather than concentration. The field distribution splits evenly at 0.5 / 0.5 between robotics / robot learning (vision-language-action models) and computer vision (LiDAR scene understanding for autonomous driving). The two 2026 papers address separate problems — subtask-guided manipulation with FOCAL-VLA versus hallucination evaluation of LiDAR language models — so current evidence shows a deliberate two-track agenda rather than a single converged line.
Collaboration58/100
Scores 58. The two cataloged papers list 17 distinct coauthors, indicating team-based work typical of robotics and vision research. Recurring collaborators in the supplied record include Di Wen, Zhiyuan Gao, Yanxiang Zhan, Mohammad Khoshnazar, Jeroen Schäfer, Michael Beetz, Runyi Yang, Murat Akkoyun, Ruiping Liu, Yufan Chen, and Junwei Zheng. Peng is last-listed author on the LiDAR benchmark paper and a middle author on FOCAL-VLA, both non-corresponding, suggesting supervisory and contributing roles within larger groups.
Activity—
Not yet scored: recent-publication frequency and year-by-year activity data are not available yet. This axis would measure publishing cadence over time, such as papers per recent year and the continuity of research activity across the career timeline.
Reproducibility78/100
Scores 78, the mean reproducibility readiness across 2 evaluated papers. The LiDAR-Hallu benchmark paper rates high: Evolve notes a rigorous protocol exposing answer bias and relation blindness in two B4DL LiDAR-LLMs, tempered only by the single model family tested. FOCAL-VLA rates medium: gains on RoboCasa and real-robot tasks are solid but statistics are thin. Together they indicate generally reusable artifacts with room for stronger evaluation.
Research summary
Kunyu Peng is a researcher whose recent recorded work sits at the intersection of robotics and computer vision, split evenly between vision-language-action (VLA) models for robot manipulation and LiDAR-based scene understanding for autonomous driving. The Evolve catalog holds two 2026 arXiv papers. In the LiDAR line, Peng is the last-listed author on "Do LiDAR Language Models Really Understand Spatio-temporal Relationships?", which introduces the LiDAR-Hallu benchmark exposing answer bias and relation blindness in two B4DL LiDAR-LLMs evaluated on nuScenes; Evolve rates its reproducibility readiness as high while noting only one model family was tested. In the robotics line, Peng appears as a middle-position author on FOCAL-VLA, which distills VGGT geometry and Track4World motion features into π0.5 for subtask-guided manipulation, with reported gains on RoboCasa and real-robot tasks but thin statistics; reproducibility readiness is medium. Across the two papers Evolve's reproducibility mean is 78 (2 papers). Collaboration data shows 17 distinct coauthors, with recurring partners including Di Wen, Zhiyuan Gao, Yanxiang Zhan, Michael Beetz, Ruiping Liu, and Junwei Zheng; the collaboration score is 58. The evenly split field distribution yields a focus score of 0. Output, impact, and activity metrics are not yet available, and institution and country are not recorded in the source data.
Collaboration pattern
The record shows 17 distinct coauthors across the two 2026 papers, consistent with team-based robotics and computer vision research. Frequently recurring names include Di Wen (appearing twice in the source list), Zhiyuan Gao, Yanxiang Zhan, Mohammad Khoshnazar, Jeroen Schäfer, Michael Beetz, Runyi Yang, Murat Akkoyun, Ruiping Liu, Yufan Chen, and Junwei Zheng. Peng holds a last-listed, senior-style author slot on the LiDAR benchmark paper and a middle slot on FOCAL-VLA; neither paper lists Peng as corresponding author.
Research keywords
Highlights
LiDAR-Hallu: a benchmark that catches LiDAR-LLM hallucination
FOCAL-VLA: distilling geometry and motion into manipulation policies
Timeline
- 2026Last-listed author on the arXiv study "Do LiDAR Language Models Really Understand Spatio-temporal Relationships?", introducing the LiDAR-Hallu benchmark for two B4DL LiDAR-LLMs on nuScenes.
- 2026Middle-position author on "FOCAL-VLA: Subtask-Guided Geometry Distillation and Implicit World Modeling for Vision-Language-Action Models", distilling VGGT geometry and Track4World motion into π0.5.
- 2026Catalog record shows work split evenly between vision-language-action robot learning and LiDAR scene understanding for autonomous driving, with 17 distinct coauthors across the two papers.
Papers in the Evolve catalog
Frequent co-authors
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 by Evolve from OpenAlex-based public bibliographic data; no institution, country, or ORCID is recorded.
- The name "Kunyu Peng" is resolved from public bibliographic data; homonyms in robotics and computer vision may exist, so identity should be verified before reuse.
- Output, impact, and activity dimensions lack underlying citation and volume metrics and are left unscored.
- The reproducibility mean of 78 rests on only 2 Evolve-reviewed papers, both from 2026; earlier work is outside this catalog.