I am a VLA Research Team Lead at AIRoA, where I focus on building robust and generalized vision-language-action models. I obtained my Ph.D. degree at UC San Diego in the CSE department under the supervision of Xiaolong Wang. Previously, I was a Research Scientist at NVIDIA Research and a Student Researcher at Google DeepMind, hosted by Kuang-Huei Lee. In the past, I've also worked with Hao Su at UCSD and Masashi Sugiyama at RIKEN-AIP.
I received my B.S. in Electrical Engineering and an M.S. in Computer Science from National Taiwan University. My research interests lie in the fields of reinforcement learning, robotics, and computer vision. Specifically, I am devoted to developing innovative methods for real-world applications. My primary focus is on building robust and generalized vision-language-action models for embodied AI.
Please see my CV for more information. If you would like to know more about my research, please contact me via email, kris.wu [at] airoa.org.
S2 keeps the original instruction as a stable goal and adds local guidance for the current step (Specify More), and puts an explicit budget on the visual tokens the policy may use, learning keep masks from action prediction alone, with no mask labels (See Less). One S2 policy covers 300+ tasks of the YUBI dataset, and across 8 real-robot tasks on TX-G2 and HSR it raises mean subtask success from 54.2% to 79.0% over π0.5.
[project page] [arXiv] In Conference on Robot Learning (CoRL), 2026