|
Zhiyan Li Hi 👋 I am Zhiyan Li, a senior undergraduate in ACM Honors Class of Shanghai Jiao Tong University, majoring in Computer Science. Currently, I'm a visiting student at Harvard University, supervised by Prof. Yilun Du and Dr. Mengyu Wang, and actively working with Weirui Ye and Kaichen Zhou at MIT. Previously, I worked at MVIG-RHOS lab in Shanghai Jiao Tong University, co-advised by Prof. Cewu Lu and Prof. Yonglu Li. My research interests include multimodal perception in robot learning (especially tactile modality), lifelong learning for robotics and test-time adaptation using steering policies and agentic methods. I am seeking a PhD position in Fall 2027. |
|
News
|
Research and Projects |
GEMA: Generative Memory for Robotic Manipulation
Zhiyan Li, Pembe Gizem Ozdil, Ben Zandonati, Ruojin Cai, Hang Le, Tomás Lozano-Pérez, Leslie Pack Kaelbling, Weirui Ye†, Yilun Du†Under CoRL Workshop Review · †Equal supervision |
SteerAct: Temporally Gated Single-Feature Steering of Frozen Action Models
Kaichen Zhou*, Zhiyan Li*, Chongkai Gao, Yuhan Wang, Wenlong Huang, Maggie Wang, Fangneng Zhan, Manling Li, Qianqian Wang, Mac Schwager, Zhuang Liu, Li Fei-Fei, Mengyu Wang†, Paul Pu Liang†Under ICLR Review · * Equal contribution; †Equal supervision |
VISTA: A Multimodal Egocentric Benchmark for Goal-Oriented Perception and Assistance to Blind and Low-Vision Users
Noé Mosca*, Mohamed Rayan Barhdadi*, Muhammad Hamza*, Joshua Yeh, Suryodaya Bikram Shahi, Zhiyan Li, Selim-Antoine Lali, Weiyue Li, HASAN KURBAN, Erchin Serpedin, Gang Luo, Yue Zhang, Minghan Li, Mengyu Wang Under ICLR Review · * Equal contribution |
|
TactiLDM for VLA: Learning Unified Visual-Text-Tactile Representations to Empower Vision-Language-Action Policies Zhiyan Li*, Hantao Jiang* web / paper / code Developed a tactile encoder within a latent diffusion framework for visual-text-tactile robot learning, improving downstream linear-probing performance by approximately 20% and increasing π0 policy success rate fourfold. |
|
Decoformer: Enhancing CrossFormer with DLinear for Long-Term Forecasting Core researcher · May 2025 code / project paper Replaced CrossFormer’s cross-time attention stage with DLinear decomposition for long-term forecasting, achieving 20% higher accuracy and 1.25× faster inference. |
|
MxCompiler Core researcher · Oct. 2024 code Implemented a Java compiler for the Mx language with semantic analysis, RISC-V code generation, and optimization passes including Mem2Reg, dead code elimination, sparse conditional constant propagation, and global-to-local variable promotion. |
|
RISC-V CPU Core researcher · Oct. 2024 code / simulator Designed a five-stage RV32IC processor in Verilog with Tomasulo-based out-of-order execution, and validated it through simulation and FPGA deployment. |
Education |
||
|
||
|
||
|
Awards and Scholarships
|
|
Credits to Jon Barron for the website design. |