Yuhang Zheng (郑宇航)

I am currently a research intern in Lightillusions and EncoSmart, focusing on 3D perception and robotic manipulation. And I am fortunate to work closely with Dr. Xiaoxiao Long, Dr. Chao Yang and Prof. Ping Tan recently. Previously, I was a research intern at Tsinghua AIR, where I started studying robotic perception.

I am an active member of AnySyn3D, a non-profit research interest group comprising individuals with a strong interest in exploring research problems and cutting-edge technologies in any topics of 3D.

I am actively looking for the PhD opportunity in Fall 2025.

Email  /  Google Scholar  /  GitHub  / 

My long-term research goal is to build agents that efficiently understand the physical world. These agents will then apply their knowledge to interact with the physical world, fostering adaptability and enabling continuous skill acquisition. So I am up for anything related research and currently focusing on {3D Vision, Robotic Manipulation} for Embodied AI:

We introduced GaussianGrasper, a robot grasping system implemented by a 3D Gaussian field endowed with open-vocabulary semantics and accurate geometry that is capable of rapid updates to support open-world robotic grasping guided by language.

We presented a simple but effective Adaptive Surface Normal (ASN) constraint to capture reliable geometric context, utilized to jointly estimate depth and surface normal with high quality.

We presented a new lift system with high lift and aerodynamic efficiency, which effectively utilizing the high lift mechanism of hummingbirds to help improve aerodynamic performance.

We designed (1) a two-stage transformer-based network to bridge the relationship between the fine-grained edges (reflectance, illumination, depth and normal) and generic edge detection tasks and (2) a cause-aware decoder, modeling edge cause as four learnable tokens.

We introduced the new task of outdoor 3D dense captioning with TOD3Cap dataset; We proposed TOD3Cap network, leveraging the BEV representation to encode sparse outdoor scenes, and combine Relation Q-Former with LLaMA-Adapter to dense captioning in the open-world.

We presented MonoOcc, a high-performance and efficient framework for monocular semantic occupancy prediction. We (1) propose an auxiliary semantic loss as supervision and an image-conditioned cross-attention module to refine voxel feature, and (2) employ a distillation module to transfer richer knowledge.

We presented 3D Implicit Transporter, a self-supervised method to discover temporally correspondent 3D keypoints from point cloud sequences. Extensive evaluations show that our keypoints are temporally consistent for both rigid and nonrigid object categories.

We presented a new interactive trajectory prediction dataset named INT2, which is short for INTeractive trajectory prediction at INTersections with high quality, large scale and rich information.

We developed an intelligent desktop operating robot designed to assist humans in their daily lives by comprehending natural language with large language models and performing a variety of desktop-related tasks.

We proposed dense prediction fields (DPFs), a new paradigm that makes dense value predictions for point coordinate queries. An implicit neural function is used to model the DPFs, which are compatible with point-level supervision.

We presented Adapt (Action-aware Driving cAPtion Transformer), a new end-to-end transformer-based framework for generating action narration and reasoning for self-driving vehicle.

We presented STEPS, the first method that jointly learns a nighttime image enhancer and a depth estimator with a self-supervised manner. And a newly proposed uncertain pixel masking strategy is used to tightly entangle these two task.

Outside of research, I enjoy playing basketball🏀, swimming🏊‍ and traveling🚢.

Website template from Jon Barron.