WristWorld: Generating Wrist-Views via 4D World Models for Robotic Manipulation

• Paper: arXiv
• Project page: wrist-world.github.io

Overview

WristWorld is a two-stage 4D world model that synthesizes realistic wrist-view videos from anchor views for robotic manipulation. It closes the gap between abundant third-person (anchor) views and scarce wrist-view recordings by explicitly modeling geometry and temporal consistency.

• Reconstruction stage: extends VGGT with a wrist head and proposes Spatial Projection Consistency (SPC) to predict geometrically consistent wrist poses and point clouds, plus wrist-view projections used as condition maps.
• Generation stage: a diffusion transformer (DiT) synthesizes wrist-view videos conditioned on wrist-view projections and CLIP-encoded anchor-view semantics, producing temporally coherent and geometrically faithful sequences.

Results on Droid, Calvin, and Franka Panda show state-of-the-art wrist-view video generation and improved VLA performance.

Repository layout

• Stage 1 — Reconstruction (VGGT + Wrist head): stage1_reconstruction/
  • Installation, training and commands are documented in stage1_reconstruction/README.md.
• Stage 2 — Generation (Video DiT): stage2_generation/
  • Environment and usage are documented in stage2_generation/README.md.

Installation

• For the reconstruction stage, follow the official VGGT training branch instructions to set up dependencies and environment: https://github.com/facebookresearch/vggt/tree/training
• For the generation stage, see stage2_generation/requirements.txt and stage2_generation/README.md.

Quick start

• Reconstruction: see stage1_reconstruction/README.md for environment setup and training commands (multiple configs provided for Droid/Calvin/RealBot variants).
• Generation: see stage2_generation/README.md for preparing condition maps and running the wrist-view video generator.

Checklist

• Open-sourced inference code
• Released weights
• Open-sourced training code

Acknowledgements

This repository builds upon and is inspired by VGGT and related geometry-grounded vision works. We thank the community for open-sourcing their research and tools.

Citation

If you find WristWorld useful, please cite our paper:

@article{qian2025wristworld,
  title   = {WristWorld: Generating Wrist-Views via 4D World Models for Robotic Manipulation},
  author  = {Qian, Zezhong and Chi, Xiaowei and Li, Yuming and Wang, Shizun and Qin, Zhiyuan and Ju, Xiaozhu and Han, Sirui and Zhang, Shanghang},
  journal = {arXiv preprint arXiv:2510.07313},
  year    = {2025}
}