Zitai Huang

Tongji University · China

Unclaimed · AI-structured profileOpenAlex

Robot-learning researcher at Tongji University extending World Action Models for vision-language-driven manipulation

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-15) 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—

Output scoring is not yet available: beyond 11 recent works and an h-index of 1, the current OpenAlex snapshot lacks the full publication record — total counts, venue types and complete per-year production — needed to compute this dimension.

Impact—

Impact cannot yet be scored: citation-based measures (citation counts, h-index trajectory, i10-index) are insufficient — the record shows only h-index 1 and i10-index 0, reflecting an early citation-accumulation stage for the 2026 preprints.

Focus0/100

OpenAlex splits this profile evenly across three robotics fields — robot manipulation/embodied AI, vision-language-action models, and machine learning for robotics, each at one third — producing the reported focus score of 0. Topic counts reinforce a robotics-learning core (Reinforcement Learning in Robotics: 7; Robot Manipulation and Learning: 7; Multimodal ML Applications: 5), with small peripheral clusters in higher education research (2) and AI planning (2) that may reflect namesake disambiguation noise.

Collaboration60/100

The collaboration score is 60, with 19 distinct coauthors across 11 recent works. Recurring partners include Taiyi Su, Tianbin Liu, Jian Zhu, Jianjun Zhang, Chong Ma, Yi Xu, Weiyi Lu and Hanli Wang. Across the three catalog papers, Huang is first author once (MaskHarness-WAM) and a middle author twice, with no corresponding-author role yet — a team-based pattern consistent with an early-career contributor inside a robotics-learning group.

Activity72/100

Activity scores 72, driven by 11 works in the last three years, all falling in 2026 in the counts-by-year series. Three are the 2026 arXiv preprints in the Evolve catalog (JEPA-WAM, MaskHarness-WAM, CSWAM), indicating a concentrated, preprint-heavy burst around world action models. Earlier records are sparse — two works in 2015 and one in 2016 — so current activity sits far above the long-run baseline.

Reproducibility32/100

Reproducibility scores 32, the mean across 3 assessed papers. Evolve's reviews flag missing artifacts: JEPA-WAM and MaskHarness-WAM have low repro readiness (no code; no task-agnostic control or significance tests; no component ablations), while CSWAM is medium — reported gains of +35/42.5 percentage points but no error bars and missing robustness baselines. Releasing code and adding statistical testing would materially raise this axis.

Research summary

Zitai Huang is a robotics and machine-learning researcher at Tongji University whose recent work centers on World Action Models (WAMs) for vision-language-driven robot manipulation. OpenAlex records 11 works with 19 distinct coauthors in the last three years, and the Evolve catalog contains three 2026 arXiv preprints. MaskHarness-WAM, with Huang as first author, re-grounds target masks per subtask for long-horizon manipulation of identical objects; CSWAM adds a V-JEPA semantic expert for out-of-distribution generalization (reported +35/42.5 percentage-point gains); JEPA-WAM injects text-to-image-generated visual instructions encoded as V-JEPA latents. Evolve's reviews note consistent real-robot gains alongside reproducibility gaps — no code, no significance tests, missing ablations — yielding a reproducibility score of 32 across three assessed papers. Bibliometric signals are early-stage: h-index 1, i10-index 0, and an even three-way field split across robot manipulation/embodied AI, vision-language-action models, and machine learning for robotics. Topic counts concentrate on reinforcement learning in robotics (7) and robot manipulation and learning (7), with a multimodal machine-learning strand (5). The identity is name-resolved from OpenAlex records without an ORCID, so the sparse 2015–2016 entries and small peripheral topic clusters (higher education research, AI planning) may include namesake noise.

Collaboration pattern

OpenAlex records 19 distinct coauthors across 11 recent works, with a stable recurring core: Taiyi Su, Tianbin Liu, Jian Zhu, Jianjun Zhang, Chong Ma, Yi Xu, Weiyi Lu and Hanli Wang. Across the three catalog papers Huang is first author once (MaskHarness-WAM) and middle author twice, never the corresponding author — a team-based pattern typical of an early-career researcher contributing modules such as semantic representations, visual instruction generation and instance grounding to a shared world-action-model program at Tongji University.

Research keywords

Robot manipulationWorld Action ModelsVision-language-action modelsInstruction followingLong-horizon manipulationInstance groundingV-JEPA latent representationsOut-of-distribution generalizationImitation learningText-to-image augmentationReinforcement learning in roboticsMultimodal machine learning

Highlights

First-author MaskHarness-WAM: instance-grounded harnessing for long-horizon manipulation
As first author, Huang proposes a VLM-gated harness that re-grounds target masks for each subtask, enabling sequential manipulation of identical objects; Evolve notes strong real-robot gains but flags the absence of component ablations.
CSWAM: causal semantic representations for OOD-robust manipulation
The paper adds a V-JEPA semantic expert to FastWAM to improve out-of-distribution generalization in imitation-learned manipulation, reporting +35/42.5 percentage-point gains; Evolve notes no error bars and missing key robustness baselines.
JEPA-WAM: generated visual instructions wired into world action models
JEPA-WAM connects text-to-image-generated visual instructions, encoded as V-JEPA latents, to world action models, reporting strong real-robot wins over Fast-WAM; Evolve flags no task-agnostic control, no significance tests and no released code.

Timeline

2015OpenAlex records 2 works under this identity — an early, sparse footprint that may reflect namesake disambiguation noise.
2016A single further work is indexed before a long gap in the record.
2026An intense publication year: 11 works recorded in the last-three-year window, all falling in 2026.
2026First-author arXiv preprint MaskHarness-WAM introduces instance-grounded harnessing for long-horizon robot manipulation.
2026Co-authored CSWAM and JEPA-WAM extend world action models with causal semantic representations and generated visual instructions.

Papers in the Evolve catalog

Frequent co-authors

Taiyi Su×3 Tianbin Liu×3 Jian Zhu×2 Jianjun Zhang×2 Chong Ma×2 Yi Xu×2 Weiyi Lu×1 Hanli Wang×1 Tianbin Liu×1 Taiyi Su×1 Zitai Huang×1 Jian Zhu×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.

Identity is name-resolved from OpenAlex bibliographic records; no ORCID is linked, so some records — especially the sparse 2015–2016 entries and small off-topic clusters — may include namesake noise.
All bibliometrics (h-index, counts by year, topics, coauthor counts) come from OpenAlex only; citation coverage for recent arXiv preprints is likely incomplete.
Output and impact dimensions are unscored because the underlying production and citation data is insufficient in the current snapshot.
All three catalog papers are 2026 arXiv preprints; peer-reviewed venue status and final published versions are not captured in this snapshot.