license: apache-2.0
datasets:
- Cognition2ActionLab/eai_real_world
language:
- en
base_model:
- Qwen/Qwen3-VL-2B-Instruct
pipeline_tag: robotics
tags:
- VLA
- Robotics
- Humanoid
HEX Model Checkpoint
This repository hosts the pretrained checkpoint of HEX: Humanoid-Aligned Experts for Cross-Embodiment Whole-Body Manipulation.
HEX is a humanoid whole-body Vision-Language-Action model for cross-embodiment manipulation. The released checkpoint contains the pretrained model weights and configuration files required for model inference.
Repository Contents
| Model | Params | Description |
|---|---|---|
| HEX | 2.4B | Pretrained HEX checkpoint for humanoid whole-body manipulation |
This repository contains the pretrained HEX checkpoint files, including model weights and the corresponding configuration files.
Download
Please follow the instructions in the HEX GitHub repository to download and use the checkpoint.
You may also download the checkpoint directly with the Hugging Face CLI:
huggingface-cli download Cognition2ActionLab/HEX-model \
--local-dir /path/to/save/HEX-model
Or with Python:
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="Cognition2ActionLab/HEX-model",
repo_type="model",
local_dir="/path/to/save/HEX-model",
)
Base Model Requirement
HEX uses Qwen3-VL as the base vision-language model. Please download the Qwen3-VL model separately before running inference.
Dataset
The processed HEX dataset is hosted in Cognition2ActionLab/eai_real_world.
License
This checkpoint is released under the Apache-2.0 License.
Citation
If you use this checkpoint, please cite our work:
@article{bai2026hex,
title={HEX: Humanoid-Aligned Experts for Cross-Embodiment Whole-Body Manipulation},
author={Bai, Shuanghao and Li, Meng and Lv, Xinyuan and Wang, Jiawei and Wang, Xinhua and Liao, Fei and Hou, Chengkai and Gu, Langzhe and Zhou, Wanqi and Wu, Kun and others},
journal={arXiv preprint arXiv:2604.07993},
year={2026}
}