Reinforcement learning researcher at UC Berkeley working across embodied AI and multi-agent AI safety
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
15
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 is not yet scored: this axis would measure total and per-year publication volume against field norms, but the snapshot lacks a verified full corpus with normalized venue and authorship metadata, so no estimate is made.
Impact—
Impact is not yet scored: this axis would measure citation-based influence, such as citations per paper and field-weighted citation percentiles, but the citation graph behind the snapshot is incomplete, so no score is reported.
Focus0/100
The focus score is 0, a direct readout of the OpenAlex field distribution: robotics/embodied AI and computational social science/multi-agent AI safety each account for exactly 0.5. Topic counts reinforce the breadth — Reinforcement Learning in Robotics (33), Adaptive Dynamic Programming Control (16), Robot Manipulation and Learning (12), Robotic Path Planning (10), and NLP Techniques (9) — so the record spans two equal tracks rather than one narrow line.
Collaboration66/100
Collaboration scores 66 from raw co-authorship data: 27 distinct coauthors in the last three years and a recurring circle of 12, including Muning Wen, Weinan Zhang, Kun Shao, Xin Wang, and Zhi Wang. On the two reviewed 2026 papers Gu appears in middle and last author positions, neither as corresponding author, indicating substantive contribution within larger team-led projects rather than sole leadership.
Activity100/100
Activity is the strongest axis at 100: the indexed record shows 42 works in the last three years, and yearly counts rise steadily from 2021 onward — 3 (2021), 5 (2022), 8 (2023), 11 (2024), 16 (2025) — with 15 already logged in 2026. This is a steep, sustained acceleration in publication rate.
Reproducibility55/100
Reproducibility scores 55 (reproMean) across 2 repro-scored papers. Both Evolve-reviewed works carry a medium readiness rating: HarnessPAI reports large simulated gains (LIBERO-PRO +61.6) but shows per-task fitting to test perturbations and thin statistics, while WolfSociety's collapse exponent is cutoff-sensitive. Limitations are documented honestly rather than hidden.
Research summary
Shangding Gu is a researcher at the University of California, Berkeley whose public record, resolved from OpenAlex bibliographic data (ORCID 0000-0002-2722-3779), centers on reinforcement learning for robots and the safety of multi-agent AI systems. The OpenAlex topic profile is led by Reinforcement Learning in Robotics (33 works), followed by Adaptive Dynamic Programming Control (16), Robot Manipulation and Learning (12), Robotic Path Planning Algorithms (10), and Natural Language Processing Techniques (9); the field distribution splits evenly, 0.5/0.5, between robotics/embodied AI and computational social science/multi-agent AI safety. Output is recent and rising: yearly counts climb from 6 works in 2020 and 3 in 2021 to 11 in 2024, 16 in 2025, and 15 already logged in 2026, with 42 works in the last three years — the basis for a maximum activity score. The reported h-index is 15 and i10-index 23. Two 2026 preprints reviewed by Evolve illustrate the agenda: HarnessPAI, a code-evolving harness built around frozen vision-language-action and world action models, and WolfSociety, a finite-size scaling study of harmful-agent spread in LLM financial agent societies. Both carry a medium reproducibility-readiness rating.
Collaboration pattern
Co-authorship is broad but team-based: 27 distinct coauthors in the last three years, with a recurring core of 12 — including Muning Wen, Weinan Zhang, Kun Shao, Xin Wang, Zhi Wang, Jian Luan, and Yang Li. On both reviewed 2026 papers Gu is neither first nor corresponding author (middle on HarnessPAI, last on WolfSociety), consistent with specialist contributor roles inside larger collaborations rather than sole-lead projects.
Research keywords
Highlights
HarnessPAI: an evolving code harness lifts frozen VLAs in simulation
WolfSociety: a scaling law for collapse in LLM financial agent societies
A dual-track agenda: embodied AI and multi-agent safety
The two reviewed 2026 papers sit on opposite sides of an even 0.5/0.5 field split — a robotics harness for physical AI and a scaling analysis of risky agent societies — showing the same learning-and-agents toolkit applied to both embodied control and safety analysis.
Timeline
- 2020Earliest year in the indexed countsByYear record, with 6 works; output dips to 3 in 2021 before sustained growth.
- 2023Output reaches 8 works, continuing a steady climb from 3 in 2021 and 5 in 2022.
- 2025Yearly output peaks at 16 works, a roughly fivefold rise over the 2021 low.
- 202615 works already recorded in the indexed counts; the Evolve catalog reviews two preprints, HarnessPAI and WolfSociety.
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 and yearly counts come from an OpenAlex-indexed snapshot; the output and impact dimensions are unscored pending fuller corpus and citation data.
- The institution is listed as UC Berkeley while the country field reads CN; both are reported exactly as provided, without further inference.
- The reproducibility mean (55) rests on only 2 scored papers, both arXiv preprints rated medium, so the estimate is thin.
- Identity is name-resolved from public bibliographic data (OpenAlex) with ORCID 0000-0002-2722-3779; no biographical facts beyond the provided fields are asserted.