A cross-embodiment framework that transfers wheeled-humanoid data to bipedal VLA models via morphology-agnostic 6D end-effector trajectories and a heuristic-enhanced online DAgger controller.

POEM-v2: a generalizable multi-view 3D hand reconstruction model trained on large-scale multi-view datasets. It enables accurate, flexible, and occlusion-robust hand mesh recovery across arbitrary multi-view setups.

AirExo-2, a low-cost exoskeleton system for large-scale in-the-wild demonstration collection. It transforms the collected in-the-wild demonstrations into pseudo-robot demonstrations. RISE-2, a generalizable imitation policy that integrates 2D and 3D perceptions.

a bidirectionally expanded learning approach that enhances auto-regressive policies for robotic manipulation. It employs a lightweight encoder-only architecture to iteratively unfold the action sequence from an initial single frame into the target sequence in a coarse-to-fine manner with logarithmic-time inference.

A two cascaded diffusion processes for object motion generation and robot action generation under object motion guidance.

a category-level garment pose tracking framework that transforms 3D point-wise learning into 2D warping deformation learning using 3D-2D projection and 2D-3D reconstruction.

Part-Aware Diffusion Model for manipulating articulated objects in real images. It uses Abstract 3D Models for efficient object representation and dynamic feature maps to transfer appearance while generating novel views or parts.

A MLLM-based method that infuses language instructions into grasp generation; & A new language-pose dataset, CapGrasp, featuring detailed caption of grasping poses.

A 4D motion dataset focusing on bimanual object manipulation tasks involved in complex daily activities; & A three-tiered task abstraction: Object Affordance, Primitive Task, and Complex Task, to systematically organize manipulation tasks.

A full-body human motion dataset that captures text-guided desktop object rearrangement through MoCap and AR glasses; & A pipeline for generating avatar's motion of object rearrangement driven by text instruction.

Reconstructing 3D implicit surfaces with accurate, view-independent surface color by decoupling view-dependent shading from geometry. It combines a global color network and a relighting network to preserve volume rendering performance while enabling colored mesh extraction.

A single-view hand-held object reconstruction method that exploits the categorical shape prior to reconstruct the shape of intra-class objects; & A new synthetic dataset, COMIC, that contains the category-level collection of objects with diverse shape, materials, interacting poses, and viewing directions.

A multi-view hand mesh recovery (HMR) method with Transformer. It leverages the "power of points", including Basis Points Set, point's positional encoding and point-Transformer, to unify and merge information from sparsely arranged cameras.

A MANO-derived hand model that contains exquisite hand-crafted texture maps, varying in appearance and covering different kinds of blemishes, make-ups, and accessories.

A dataset that focuses on human grasp based on object's affordance. It contains two knowledge base: 1) Object affordance knowledge (Oak) and 2) Interaction knowledge (Ink).
A new model: Tink, for transferring interaction pose from one object to another.

An online data syhthesis tool for articulated hand(-object) pose estimation. An grasping systhesis method that can generate dexterous hand grasping poses for arbitrary object.

A novel contact representation (CPF) that used to imporve physical hand-object interaction. A hybrid learning-fitting framework (MIHO) that aligns the top-down pose estimation with bottom-up contact modeling.

A novel contact representation (CPF) that used to imporve physical hand-object interaction. A hybrid learning-fitting framework (MIHO) that aligns the top-down pose estimation with bottom-up contact modeling.

A hybrid inverse kinematics method for 3D body mesh recovery, combining 3D keypoint estimation and body mesh recovery. HybrIK-X extends this to model hands and faces, offering fast, accurate whole-body pose estimation.