IronMind: Scaling Humanoid Dexterous Manipulation via Camera-Space Ego-Centric Pretraining
Technical report, 2026
Learning humanoid dexterous manipulation from human egocentric video with a shared camera-space action representation.
Jiacheng Wei
Staff Research Scientist · XPENG Robotics
My work focuses on training embodied foundation models and physical agents, and building virtual worlds.
I serve as technical co-lead of the World Model team and technical lead of the Robotics Simulation team at XPENG Robotics.
Previously, I was a postdoctoral research fellow at Nanyang Technological University, Singapore, where I also received my Ph.D., advised by Prof. Guosheng Lin and Prof. Kim-Hui Yap. I interned at ByteDance with Dr. Jiashi Feng.
Selected research
* Equal contribution · † Corresponding author. Full list on Google Scholar.
15 papers
XPACE: Joint World and Action Modeling from Heterogeneous Experience
Technical report, 2026
One model learns from human and robot experience, predicts actions and video, and simulates recovery trajectories to improve real robot behavior.
AnyWorld: Factorized Egocentric World Models for Cross-Embodiment Generalization
CoRL 2026
Recompose human interactions across bodies, viewpoints, and scenes to generate robot-native experience.
IC-World: In-Context Generation for Shared World Modeling
ECCV 2026
Generate multiple views of one shared world, with consistent geometry and motion across videos.
iMontage: Unified, Versatile, Highly Dynamic Many-to-many Image Generation
CVPR 2026
Turn video-model priors into a flexible image generator that preserves context across diverse image sets.
DiVA: Enabling Interactive Digital Life Simulation via Video Models
arXiv 2026
Long-term interactive digital character simulation via anchored video continuation.
EditVerse3D: High-Quality 3D Object Editing with Region-Aware Learning
ECCV 2026
Edit a 3D object from a coarse region and a reference image while preserving the rest.
Adaptive Piecewise Distillation for Efficient LiDAR Data Generation
AAAI 2026
Generating realistic LiDAR point clouds in four sampling steps through adaptive distillation.
Puppeteer: Rig and Animate Your 3D Models
NeurIPS 2025
Bring static 3D assets to life with automatic skeletons, skinning, and video-guided animation.
PhyCustom: Towards Realistic Physical Customization in Text-to-Image Generation
arXiv 2025
Customizing physical concepts in text-to-image generation through isometric and concept-decoupling regularization.
CADCrafter: Generating Computer-Aided Design Models from Unconstrained Images
CVPR 2025
From an everyday image to an editable, parametric CAD model.
OccLE: Label-Efficient 3D Semantic Occupancy Prediction
arXiv 2025
Learning semantic 3D occupancy from images and LiDAR with only a fraction of voxel annotations.
Sync4D: Video Guided Controllable Dynamics for Physics-Based 4D Generation
arXiv 2024
Transfer motion from video into controllable, physics-based 4D content.
REACTO: Reconstructing Articulated Objects from a Single Video
CVPR 2024
Recover the shape and articulation of moving objects from one casual video.
Prim2Room: Layout-Controllable Room Mesh Generation from Primitives
arXiv 2024
Build room meshes with spatial control through simple geometric primitives.