Yufeng Li 李雨峰

Ph.D. Student @ Shanghai Jiao Tong University

I am a Ph.D. student at Shanghai Jiao Tong University, advised by Prof. Junchi Yan. I am also jointly trained at Beijing Zhongguancun College, advised by Researcher Kai Chen.

My research focuses on embodied intelligence for precise and safety-critical robotic systems. I am interested in scientific robotics, dexterous manipulation, vision-language-action models, safe reinforcement learning, optimal transport, and generative policy learning.

Recent projects study how robots can operate in chemical-laboratory environments, how VLA models can specialize task-relevant action factors, and how optimal-flow views can support safer reinforcement learning.

Email GitHub CV

Updates

[Aug. 2026]	SynMatAI AI4S Future ScienceSkills Hackathon - Champion (1st of 82 teams).
[Jun. 2026]	SafeLab, ICML 2026.
[May 2026]	GuidedVLA, RSS 2026.
[Apr. 2025]	Optimal Flow Transport and its Entropic Regularization, ICLR 2025.
[Oct. 2024]	REFRAME, ECCV 2024.

Selected Publications

SafeLab: An Interactive High-Fidelity Benchmark for Embodied Safety in Scientific Robotics

F. Bai*, Y. Li*, R. Wu, P. Wang, Y. Wang, B. H. Zhu, Y. Wang, T. Chou, J. Gao, R. Zhu, Y. Wen, Y. Yang, Y. Chen, et al.

International Conference on Machine Learning (ICML), 2026

High-fidelity chemical-lab benchmark with LLM-programmed task synthesis, automated expert demonstrations, and interactive RL post-training.

Paper Demo Video Talk Video

Optimal Flow Transport and its Entropic Regularization: a GPU-friendly Matrix Iterative Algorithm for Flow Balance Satisfaction

L. Shi, Y. Li, K. Zeng, Y. Tu, J. Yan, et al.

International Conference on Learning Representations (ICLR), 2025

A matrix-iterative optimal transport algorithm for satisfying node and edge flow balance constraints in graph-structured problems.

Paper Code

GuidedVLA: Specifying Task-Relevant Factors via Plug-and-Play Action Attention Specialization

X. Jia, B. Yang, Z. Ge, X. Nie, Y. Zhou, C. Fan, Y. Li, Y. Chai, C. Jing, Z. Liang, Q. Bu, H. Cao, C. Wu, Q. Li, Z. Yang, C. Zhang, H. Li, Z. Wu, J. Yan, Y.-G. Jiang

Robotics: Science and Systems (RSS), 2026

Decouples object pointing, spatial geometry, and timing skill factors by specializing action-decoder attention heads.

Paper Homepage

REFRAME: Reflective Surface Real-Time Rendering for Mobile Devices

C. Ji, Y. Li, Y. Liao, et al.

European Conference on Computer Vision (ECCV), 2024

Real-time novel-view synthesis pipeline for reflective surfaces on mobile devices.

Paper Code Homepage

Efficient Packaging Line Object Counting by Cross-frame Association with Wavelet Convolutions and Trajectory Compensation

L. Wei, Y. Zhu, Y. Li, M. Qian, X. Zuo, B. Chen, S. Liang, Z. Lin, J. Yan, et al.

IEEE Access, 2025

Wavelet-convolution enhanced detection and cross-frame association for packaging line object counting.

Paper

Additional Publications & Manuscripts

Towards Practical Large-scale Dynamical Heterogeneous Graph Embedding: Cold-start Resilient Recommendation M. Long, J. Liu, Y. Li, H. Xiong, J. Yan, K. Wang, Y. Cao, J. Ding IEEE Transactions on Knowledge and Data Engineering (TKDE), 2025.
GEARS: Seeing Geometry, Diffusing Actions for Zero-Shot Sim-to-Real Dexterous Manipulation F. Bai, T. Chou, Y. Li, Y. Li, Y. Wang, Y. Mao, H. Zhang, Y. Wang, R. Zhu, Y. Wen, Y. Yang, Y. Chen, et al. Submitted to AAAI, 2026.
SafeTransport: A Unified Capacity-Flow View of Safe Reinforcement Learning Y. Li, F. Bai, T. Chou, Q. Li, J. Gao, J. Yan Submitted to NeurIPS, 2026.
Rsync: Reward-Manifold Driven Adaptive Synchronization for Sample-Efficient Real-World Reinforcement Learning Y. Cai, F. Bai, T. Yu, Y. Pu, Y. Mao, Y. Wang, T. Chou, Y. Li, Y. Li, Y. Yang, H. Zhang, M. Li, Y. Chen, et al. Submitted to AAAI, 2026.
Remembering What Matters: From Markovian to Subtask-Causal Memory in VLA Policies Peishuo Wang, Fengshuo Bai, Yufeng Li, Tawei Chou, Yinda Xu, Ying Wen, Yaodong Yang, Chen Gao, Zhenzhe Zheng, Fan Wu Submitted to NeurIPS, 2026. Current review scores: 5 / 4 / 4.
Beyond Polarity Counting: A Polarity-Entropy Framework for Label-Free Neural SAT Phase Guidance Y. Li*, X. Zheng*, J. Yan Submitted to NeurIPS, 2026. Current review scores: 5 / 5 / 5 / 4.

Experience

Research Intern, Embodied Intelligence, Beijing PsiBot Technology Sep 2025 - Present
Built SafeLab with LLM-based task synthesis and automated expert demonstrations across 6,000+ trajectories, 63 assets, and 64 tasks.
Recommendation System Algorithm Research Intern, Huawei Mar 2025 - Sep 2025
Built a deployable two-stage dynamic graph embedding system for real-time incremental updates on billion-scale graphs.

Education and Honors

Shanghai Jiao Tong University, Ph.D. student, 2024 - 2029. Research direction: embodied intelligence, optimal transport, and generative models.
University of Electronic Science and Technology of China, B.Eng. in Software Engineering, 2020 - 2024. Professional ranking: 1/202.
Champion, SynMatAI AI4S Future ScienceSkills Hackathon, 2026 (1st of 82 teams); National Scholarship in 2021 and 2023; National Second Prize in Computer Design Contest; Outstanding Graduate of Sichuan Province.

Selected Project

Hackathon Champion | Agentic Science Skill

Global Geochemical Atlas Skill

One prompt turns a research question into a trustworthy, traceable, and interactive atlas of the temporal evolution of global geochemical element distributions.

Global Geochemical Atlas Skill translates heterogeneous requests from researchers into one unified, Agentic-Native workflow for trustworthy, interactive, traceable, and time-aware geochemical research.

Homepage GitHub Slides Zhihu Xiaohongshu WeChat

Local-first AI workflow system

ShrimpFlow

A programmer workflow system that observes terminal, git, and AI-agent interaction traces, models a developer's workflow and cognitive preferences, and exports executable behavior specifications such as ClawProfile / CLAUDE.md for personalized AI coding agents.

Shadow layer: terminal, git, agent-session, and environment event collection.
Brain layer: workflow, preference, decision, and temporal pattern modeling.
Autopilot layer: portable behavior specs for downstream coding agents.

GitHub Live workflow MP4 demo Bilibili demo

Global Reach

Visitors Worldwide

Approximate visitor locations and total pageviews.

World map of approximate visitor locations

Locations are approximate; no precise personal location is displayed. Visitor statistics