I am a Ph.D. student at Texas A&M University advised by Prof. Zhiwen Fan, and a Senior Research Engineer at Google DeepMind working on Gemini post-training and safety alignment.
Previously, I received my M.Sc. from Carnegie Mellon University and my B.Sc. from the University of Edinburgh, where I was advised by Prof. Hakan Bilen.
MoRE removes unwanted behavior modes from behavior-cloned policies, such as handing over a knife blade-first, with a short uncloning step and zero inference-time overhead.
A 500-hour dataset pairing egocentric video with full-hand tactile signals; scaling it raises robot task success from 22.5% to 57.5% and improves zero-shot touch prediction on unseen sensors.
Lulin Liu*, Nuo Chen*, Yan Wang, Bangya Liu, Wenyan Cong, Hezhen Hu, Boris Ivanovic, Hao Wang, Ziyao Zeng, Xinyu Gong, Yang Zhou, Zixiang Xiong, Dilin Wang, Zhangyang Wang, Weisong Shi, Ruohan Zhang, Marco Pavone, , Zhiwen Fan
An open-source generative data engine that turns long-tail driving videos, even single dash-cam clips, into pose-grounded multi-view data for VLA policy learning.
Huiwen Han*, Nuo Chen*, Ziyao Zeng*, Dayou Li, Hao Wang, Amir Hossein Rahmati, Xiaoning Qian, Bangya Liu, Zhiwen Fan
Rebuilds TouchScale glove recordings in simulation with touch-driven contact forces; policies trained on these replays grip at 0.99× the recorded force, versus 0.31× for pose imitation.
A live attack-and-defense CTF benchmark: two agents exploit each other's service while patching their own over multiple rounds, with a latent-ability model placing offense, defense and difficulty on one scale.
Dream2Fix synthesizes photorealistic failures from successful demonstrations with a generative world model, producing 120K+ paired samples for learning executable recovery.