Robot learning and multimodal AI evaluation researcher at Georgia Tech; VLA post-training and omni-model benchmarking
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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: the current snapshot lacks reliable publication-volume and venue data for this researcher, so output cannot be computed. This axis would measure total indexed works, venue tier, and first-author share across the career record.
Impact—
Not yet scored: citation-based impact cannot be computed from the current snapshot, and its h-index of 0 likely reflects an early or incompletely disambiguated OpenAlex record. This axis would measure citations, citation velocity, and influence on subsequent research.
Focus0/100
Scores 0, the lowest band. The field distribution splits evenly at 0.5 for computer vision (generative models / multimodal AI evaluation) and 0.5 for robot learning (vision-language-action models, robot manipulation), while the topic list spans five areas: reinforcement learning in robotics (2), robot manipulation and learning, GANs and image synthesis, domain adaptation and few-shot learning, and stochastic gradient optimization.
Collaboration63/100
Scores 63 on a 0-100 scale. The snapshot records 23 distinct coauthors, and Evolve's frequent-collaborator list includes Haoyu Zhao, Tianyu Deng, Ziqin Xu, Xudong Wang, and Shuicheng Yan. Both reviewed 2026 papers place this researcher in a middle-author, non-corresponding role, consistent with contributing inside larger robotics and multimodal-evaluation teams rather than leading solo or small-group efforts.
Activity40/100
Scores 40. OpenAlex indexes 3 works over the last three years, and counts-by-year places all of them in 2026, indicating a very recent publication start with no indexed output in earlier years. The moderate score reflects this thin, short record rather than any slowdown: both 2026 arXiv preprints have already entered Evolve's review pipeline.
Reproducibility20/100
Scores 20, matching the reproducibility mean of 20 computed across 2 reviewed papers. Evolve rates readiness as low for both: the HIL-UMI study reports clear SFT gains but no variance data, and the MiniMax-H3 evaluation covers a single model, so the reviewed evidence does not yet support confident independent replication of either result.
Research summary
Zihao Zhao is an early-stage researcher at Georgia Institute of Technology whose OpenAlex-indexed work sits at the intersection of robot learning and computer vision, with the field distribution split evenly at 0.5 between robot learning (vision-language-action models, robot manipulation) and computer vision (generative models / multimodal AI evaluation). The snapshot counts 3 works in the last three years, with counts-by-year placing all of them in 2026, and Evolve has reviewed two 2026 arXiv preprints. The first, HIL-UMI, builds a robot-free, human-in-the-loop post-training pipeline for vision-language-action models on the Universal Manipulation Interface, combining energy-score out-of-distribution detection with advantage-conditioned behavior cloning; Evolve notes clear SFT gains but no variance data. The second is a human-evaluated 517-case benchmark probing MiniMax-H3's physical-world reasoning, which the model scores at 41.97%, judged weak. Topic data reinforces this dual profile: reinforcement learning in robotics appears twice, alongside robot manipulation and learning, GANs and image synthesis, domain adaptation and few-shot learning, and stochastic gradient optimization. The record lists 23 distinct coauthors, and both reviewed papers carry middle-author, non-corresponding positions, consistent with a team-based early-career contributor embedded in larger groups. Citation metrics in the snapshot (h-index 0, i10-index 0) are best read as a very early or incompletely disambiguated record rather than evidence about influence.
Collaboration pattern
The record shows a wide, team-oriented collaboration footprint: 23 distinct coauthors, with Evolve flagging Haoyu Zhao, Tianyu Deng, Ziqin Xu, Xudong Wang, and Shuicheng Yan among the most frequent. Both reviewed papers are middle-author contributions on multi-author teams spanning robot learning and multimodal evaluation, and neither lists this researcher as corresponding author, consistent with an early-career contributor embedded in larger lab groups.
Research keywords
Highlights
Robot-free human-in-the-loop post-training of VLA models via UMI
517-case human evaluation probes MiniMax-H3's physical-world reasoning
Timeline
- 2026First indexed publication year: OpenAlex counts-by-year places all 3 works in 2026, with recent3yWorks also at 3.
- 2026Releases the HIL-UMI preprint on arXiv, introducing robot-free human-in-the-loop post-training of vision-language-action models on the Universal Manipulation Interface.
- 2026Publishes the MiniMax-H3 evaluation preprint on arXiv, a 517-case human benchmark finding weak physical-world reasoning at 41.97%.
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.
- Bibliometrics come from an OpenAlex-resolved record only; no ORCID is linked, and Zihao Zhao is a common name, so identity carries some disambiguation uncertainty.
- In-depth coverage is limited to the two Evolve-reviewed 2026 arXiv preprints; other outputs may exist but were not reviewed.
- Snapshot citation metrics (h-index 0, i10-index 0) likely reflect a very early or incompletely linked record, so output and impact dimensions are left unscored rather than estimated.
- Reproducibility ratings and verdicts reflect Evolve's judgments at review time on the two papers and may change as code, data, or additional results are released.