Publications

Progress Predicts Policy

Licheng Zhong, Gim Hee Lee

arXiv preprint, 2026
Project Page / arXiv / Code

Progress Predicts Policy (PPP) represents robot tasks as language-conditioned scalar fields that directly induce actions through their local geometry. Composed with a learned world model, the field generates continuous motions from gradients and discrete gripper decisions from finite differences, while also providing progress estimates for execution monitoring.

Diagnosis-Guided Recovery from Curriculum Stalls in Agentic Robot Reinforcement Learning

Hui Yang, Wei Sun, Licheng Zhong, Gim Hee Lee

arXiv preprint, 2026
arXiv

CADRE is an agentic framework for diagnosing and recovering from curriculum stalls in robot reinforcement learning. It uses fixed physical milestones to identify blocked transitions and guide targeted reward repair, verifying recovery without changing the behavioral goals. By reusing acquired skills, CADRE improves manipulation task completion while reducing training interactions..

Wear and Go: Calibration Free Sparse Inertial Motion Capture via Online Sensor Setup Estimation

Yuanshuo Tan, Licheng Zhong

arXiv preprint, 2026
Project Page / arXiv

Wear and Go enables full-body motion capture with six interchangeable IMUs worn at prescribed body sites, without a calibration pose. Its MOSAIC framework accumulates evidence from ordinary motion to recover sensor identities and mounting orientations, then aligns the measurements for pretrained pose estimators without retraining.

Grasp-Oriented Non-Prehensile Manipulation via Learning a Graspability Field

Licheng Zhong, Gim Hee Lee

European Conference on Computer Vision (ECCV), 2026
Project Page / Paper / arXiv / Code

GOMP formulates non-prehensile manipulation for grasping as optimizing graspability rather than reaching a predefined target pose. It learns a graspability field over object configurations, providing a dense training signal for closed-loop object reconfiguration, autonomous stopping, and direct transition to grasp execution without external planners or manually specified target states.

Bi-Adapt: Few-shot Bimanual Adaptaion for Novel Categories of 3D Objects via Semantic Correspondence

Jinxian Zhou*, Ruihai Wu*, Yiwei Liu, Yiwen Hou, Xunzhe Zhou, Checheng Yu, Licheng Zhong, Lin Shao

International Conference on Robotics & Automation (ICRA), 2026, Best Paper Finalist
Project Page / arXiv / Code

Bi-Adapt is a diffusion-based framework for efficient and generalizable bimanual manipulation across novel object categories and complex tasks.

Full-Body Golf Swing Kinematic Reconstruction From a Smartwatch IMU

Yuanshuo Tan, Kezhe Zhu, Xiujie Sun, Chunping Liang, Shuoyang Zhu, Chenquan Xu, Licheng Zhong, Huiming Pan, Yinri Jin, Chang Liu, Bo Xiao, Shenglong Le, Bryndan W. Lindsey, Peter B Shull

arXiv preprint
arXiv

This work presents the first method for reconstructing full-body golf swing kinematics from a single wrist-worn IMU, enabling practical swing analysis during real-world play.

Transformer-Based Full-Body Pose Estimation for Rehabilitation via RGB Camera and IMU Fusion

Yuanshuo Tan, Xinyuan He, Guoxing Liu, Licheng Zhong, Huiming Pan, Kezhe Zhu, Peter B Shull

International Conference on Body Sensor Networks (BSN), 2025, Oral Presentation
Paper

This work proposes a temporal transformer-based method that fuses images and IMU data to improve 3D human pose estimation for specialized rehabilitation exercises.

Mimicking-Bench: A Benchmark for Generalizable Humanoid-Scene Interaction Learning via Human Mimicking

Yun Liu*, Bowen Yang*, Licheng Zhong*, He Wang, Li Yi

CVPR Workshop on Humanoid Agents, 2025, Spotlight
Project Page / arXiv

A benchmark of Humanoid Scene Interaction. The first comprehensive benchmark for generalizable humanoid-scene interaction learning via human mimicking. Integrated a large-scale diverse human skill reference dataset with both synthetic and real-world human-scene interactions. Developed a general skill-learning paradigm and provide support for both pipeline-wise and modular evaluations.

Multi-view Hand Reconstruction with a Point-Embedded Transformer

Lixin Yang, Licheng Zhong, Pengxiang Zhu, Xinyu Zhan, Junxiao Kong, Jian Xu, Cewu Lu

Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2025
Paper / arXiv / Code

POEMv2 is a generalizable multi-view hand mesh reconstruction model which embeds a static basis point within the multi-view stereo space.

Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians

Licheng Zhong, Hong-Xing "Koven" Yu, Jiajun Wu, Yunzhu Li

European Conference on Computer Vision (ECCV), 2024
Project Page / Paper / arXiv / Data / Code

Spring-Gaus reconstructs the appearance, geometry, and physical dynamics properties of elastic objects from video observations. Spring-Gaus enables future prediction and simulation under different initial states and environmental parameters.

Color-NeuS: Reconstructing Neural Implicit Surfaces with Color

Licheng Zhong*, Lixin Yang*, Kailin Li, Haoyu Zhen, Mei Han, Cewu Lu

International Conference on 3D Vision (3DV), 2024
Project Page / Paper / arXiv / Data / Code

Color-NeuS focuses on mesh reconstruction with color. We remove view-dependent color while using a relighting network to maintain volume rendering performance. Mesh is extracted from the SDF network, and vertex color is derived from the global color network. We conceived a in hand object scanning task and gathered several videos for it to evaluate Color-NeuS.

CHORD: Category-level in-Hand Object Reconstruction via Shape Deformation

Kailin Li*, Lixin Yang*, Haoyu Zhen, Zenan Lin, Xinyu Zhan, Licheng Zhong, Jian Xu, Kejian Wu, Cewu Lu

International Conference on Computer Vision (ICCV), 2023
Project Page / Paper / arXiv / PyBlend

We proposed a new method CHORD which exploits the categorical shape prior for reconstructing the shape of intra-class objects. In addition, we constructed a new dataset, COMIC, of category-level hand-object interaction. COMIC encompasses a diverse collection of object instances, materials, hand interactions, and viewing directions, as illustrated.

POEM: Reconstructing Hand in a Point Embedded Multi-view Stereo

Lixin Yang, Jian Xu, Licheng Zhong, Xinyu Zhan, Zhicheng Wang, Kejian Wu, Cewu Lu

Computer Vision and Pattern Recognition (CVPR), 2023
Paper / arXiv / Code

POEM (Point Embedded Multi-view) focuses on reconstructing an articulation body with "true scale" and "accurate pose" from a series of sparsely arranged camera views. In practice, we used the example of hand. POEM explores the power of points, using a cluster of (x, y, z) coordinates with natural positional encoding to find associations in multi-view stereo.

Centralized and Decentralized Methods for Multi-robot Safe Navigation

Xinbei Wang*, Zexuan Yan*, Licheng Zhong*

Machine Learning and Intelligent Systems Engineering (MLISE), 2022
Paper

We simulated two centralized methods and two decentralized methods for multi-robot safe navigation in their respective environments.