Yinhuai Wang

Hi! I am a second-year Ph.D student at HKUST, under the supervision of Prof. Ping Tan. I hold a Master's degree from Peking University and a Bachelor's degree from Xidian University. Additionally, I have interned at IDEA Research, Unitree Robotics, and Shanghai AI Lab.

Before my postgraduate studies, I worked at DH-Robotics as one of the initial team members (the first engineer), where I led a team in the development of electronics and algorithms for robot grippers and arms.

Email  /  CV  /  Google Scholar  /  Github  /  Zhihu

My research interests lie in enabling humanoids to interact agilely and robustly with the world (with objects, scenes, humans, and other agents). Since 2023, I have pioneered the Interaction Imitation paradigm, under which I have led a series of projects including PhysHOI, SkillMimic, SkillMimic-V2, HOP-HOT, and HumanX. My long-term goal is endowing humanoids with dynamic interaction capabilities that surpass human performance.

We present HumanX, a full-stack framework that compiles human video into generalizable, real-world interaction skills for humanoids, without task-specific rewards.

By introducing novel data augmentation techniques for skill transitions and recovery patterns, we acquire robust and generalizable interaction skills from noisy and sparse HOI demonstrations.

We for the first time enable simulated humanoids to learn basketball interaction skills purely from HOI demonstrations. The learned interaction skills can be reused for complex and long-horizon tasks such as continuous scoring.

We enable physically simulated humanoids to imitate interactions from video demonstrations, without designing task-specific rewards.

We bring Range-Null space Decomposition (RND) into diffusion models, enabling diverse image restoration tasks in a zero-shot manner, without extra training or optimization.

We bring Range-Null space Decomposition (RND) into GAN-Prior based SR models to accelerate the convergence and ensure the downsampling consistency.

FreeDoM is a simple but effective training-free method generating results under control from various conditions using unconditional diffusion models.

We find that simply adding an independent LN to each layer can robustly improve the performance of vision transformers.

We fuse the generative prior and the image prior dynamically according to the degradation levels.

This cool template is stolen from Jon Barron!