FlowHOI: Flow-based Semantics-Grounded Generation of Hand-Object Interactions for Dexterous Robot Manipulation

Pretrained checkpoints for FlowHOI, a two-stage framework for hand-object interaction generation: an EgoDex grasping prior followed by GRAB or HOT3D manipulation.

Paper	arXiv 2602.13444
Project page	https://huajian-zeng.github.io/projects/flowhoi/
Code	https://github.com/huajian-zeng/flowhoi
Demo data	huajian-zeng/flowhoi-data

Files

File	Description
egodex_grasp/model000500000.pt	Shared grasping prior trained on EgoDex
grab_full/model000200000.pt	GRAB manipulation model
hot3d_full/model000200000.pt	HOT3D manipulation model with scene conditioning

Each checkpoint has an accompanying args.json configuration file.

Usage

Clone the FlowHOI repository and follow its installation instructions. Then run from the repository root:

bash scripts/download_pretrained.sh
bash scripts/download_data.sh
bash demo.sh                  # GRAB
DATASET=hot3d bash demo.sh     # HOT3D

This release provides inference and visualization with three prepared demo sequences. Rendering requires the separately obtained MANO/SMPL-X and dataset assets described in the code README.

Code license

Citation

@article{zeng2026flowhoi,
  title   = {{FlowHOI}: Flow-based Semantics-Grounded Generation of Hand-Object Interactions for Dexterous Robot Manipulation},
  author  = {Zeng, Huajian and Chen, Lingyun and Yang, Jiaqi and Zhang, Yuantai and Shi, Fan and Liu, Peidong and Zuo, Xingxing},
  journal = {arXiv preprint arXiv:2602.13444},
  year    = {2026}
}