DECO_sim

DECO: Decoupled Multimodal Diffusion Transformer for Bimanual Dexterous Manipulation with a Plugin Tactile Adapter

<p> <a href="https://arxiv.org/pdf/2602.05513"> <img src="https://img.shields.io/badge/arXiv-Paper-b31b1b?logo=arxiv"> </a> <a href="https://baai-humanoid.github.io/DECO-webpage/"> <img src="https://img.shields.io/badge/Project-Website-blue"> </a> <a href="https://huggingface.co/datasets/BAAI-Humanoid/DECO-50"> <img src="https://img.shields.io/badge/HuggingFace-Data-yellow?logo=huggingface"> </a> <a href="https://huggingface.co/BAAI-Humanoid/DECO_sim"> <img src="https://img.shields.io/badge/HuggingFace-Model-red?logo=huggingface"> </a> <a href="LICENSE"> <img src="https://img.shields.io/badge/License-Apache--2.0-green"> </a> </p>

All checkpoints use a compact model of about 80M parameters and achieve near-SOTA success rates on their respective benchmarks.

Checkpoints

Results

LIBERO

Success rates (50 trials per task):

DexJoCo

Success rates (50 trials per task); /B denotes a bimanual task:

UniVTAC

Success rates (100 trials per task):

RoboTwin

Success rates (%) on the 50 RoboTwin dual-arm tasks under the Seen and Unseen settings (100 trials per task):

Usage

Download all checkpoints:

hf download BAAI-Humanoid/DECO_sim --repo-type model --local-dir ./sim_ckpt/

Load a checkpoint into the DECO policy and evaluate it with the corresponding benchmark configuration. The training and evaluation code lives in the DECO repository:

• LIBERO, DexJoCo, and UniVTAC are on the sim branch.
• RoboTwin is on the robotwin branch.

Citation

@article{li2026deco,
  title={DECO: Decoupled Multimodal Diffusion Transformer for Bimanual Dexterous Manipulation with a Plugin Tactile Adapter},
  author={Xukun Li and Yu Sun and Lei Zhang and Bosheng Huang and Yibo Peng and Yuan Meng and Haojun Jiang and Shaoxuan Xie and Guacai Yao and Alois Knoll and Zhenshan Bing and Xinlong Wang and Zhenguo Sun},
  journal={arXiv preprint arXiv:2602.05513},
  year={2026}
}

License

Apache 2.0