Robot learning researcher at UNC Chapel Hill focused on vision-language-action policies and manipulation evaluation.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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 measures publication volume, venue quality, and productivity over time. This axis is not yet scored: the underlying data covers only two cataloged 2026 arXiv papers, with no full publication or venue record available.
Impact—
Impact would capture citation counts, scholarly influence, and uptake of HALTER and EGR beyond the catalog. No citation data is available yet, so this axis stays unscored rather than estimated.
Focus0/100
The focus axis scores 0 on the current 0-100 scale. The underlying field distribution is an exact 50/50 split between robot learning (vision-language-action models / multimodal robotics) and robotics (robot learning and manipulation evaluation). Both are closely adjacent subfields, and the two cataloged 2026 papers share vision-language-action keywords, so the even split reflects neighboring topics rather than unrelated research directions; a larger sample could shift this reading.
Collaboration46/100
Collaboration scores 46, based on 9 distinct coauthors in the cataloged record. Recurring collaborators include Daniel J. Szafir, Yue Yang, Diego Romeres, Chiori Hori, Siddarth Jain, Jing Jiang, Xinkai Jiang, and Rudolf Lioutikov. Both 2026 papers list Bertasius as a middle author rather than corresponding author, consistent with contribution inside larger team projects; the moderate score reflects the modest recorded coauthor pool, not an absence of collaboration.
Activity—
Activity tracks recent publishing cadence and year-over-year research output. The data covers only 2026 arXiv submissions, so a multi-year activity trend cannot be computed; this axis remains unscored until a longer publication history is available.
Reproducibility38/100
Reproducibility scores 38, computed as a mean over 2 papers. The manipulation-evaluation paper (HALTER) is rated low readiness: Evolve notes solid real-robot gains but thin small-N statistics. The VLA robustness paper (EGR) is rated medium: notable simulated and real gains and honest statistics, but key ablations and baselines are missing. Neither paper is currently accompanied by independently verified replication.
Research summary
Gedas Bertasius is a researcher at the University of North Carolina at Chapel Hill whose cataloged work centers on robot learning, split evenly (50/50 in Evolve's field distribution) between vision-language-action models and multimodal robotics on one side, and robot learning and manipulation evaluation on the other. Two 2026 arXiv papers appear in the Evolve catalog, in both of which Bertasius is a middle author rather than corresponding author. The first, 'From Rollout to Reset', introduces HALTER, a graph-based harness that plans atomic-skill resets with an LLM over scene graphs for autonomous long-horizon manipulation evaluation; Evolve notes solid real-robot gains but thin small-N statistics and rates reproducibility readiness low. The second, 'Sensing Which Modality Matters', proposes evidence-gated regularization (EGR), gating consistency losses by per-sensor evidence to improve vision-language-action robustness, with notable simulated and real gains and honest statistics, though key ablations and baselines are missing; readiness is medium. Across the radar, reproducibility scores 38 (mean over two papers), collaboration 46 with nine distinct coauthors, and focus 0 on the 0-100 scale; output, impact, and activity are not yet scored for lack of data.
Collaboration pattern
Bertasius's cataloged record shows 9 distinct coauthors and a collaboration score of 46. Recurring names include Daniel J. Szafir, Yue Yang, Diego Romeres, Chiori Hori, Siddarth Jain, Jing Jiang, Xinkai Jiang, and Rudolf Lioutikov. In both 2026 papers Bertasius appears as a middle author, not corresponding author, suggesting work embedded in larger teams. The recorded pool spans evaluation, VLA robustness, and related robotics topics; a fuller publication history would clarify the true network.
Research keywords
Highlights
HALTER: graph-based harness for autonomous long-horizon manipulation evaluation
Evidence-gated regularization for robust vision-language-action policies
Timeline
- 2026Publishes 'Sensing Which Modality Matters: Evidence-Gated Regularization for Robust VLA Policies' on arXiv, applying evidence-gated regularization and imitation learning on BEHAVIOR-1K for multimodal robustness.
- 2026Publishes 'From Rollout to Reset: A Graph-Based Harness for Autonomous Long-Horizon Manipulation Evaluation' on arXiv, introducing HALTER's LLM-planned scene-graph resets for real-robot evaluation.
- 2026Concentrates all cataloged output in this single year, split evenly between robot learning (VLA/multimodal robotics) and robotics (manipulation evaluation), with 9 distinct coauthors recorded.
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 from a small Evolve catalog of two 2026 arXiv papers; output, impact, and activity remain unscored for lack of data.
- The reproducibility mean of 38 covers only 2 reviewed papers; low/medium readiness reflects Evolve review judgments, not external replication.
- Citation counts, h-index, venue history, and funding information were not available and are not estimated anywhere in this profile.
- Field distribution and coauthor statistics are name-resolved from public bibliographic data (ORCID 0000-0003-1800-4790) and reflect only cataloged papers.