<h1 align="center"> Modality-Composable Diffusion Policy via Inference-Time Distribution-level Composition<br> </h1>

Accepted to <i style="color: black; display: inline;"><b>Generative Models for Robot Learning Workshop @ ICLR 2025</b></i>

PDF | arXiv | ICLR Genbot Homepage<br>

Jiahang Cao, Qiang Zhang, Hanzhong Guo, Jiaxu Wang, Hao Cheng, Renjing Xu.

HKUSTGZ, Beijing Innovation Center of Humanoid Robotics, HKU

<div align="center"> <img src="main.png" alt="dp3" width="90%"> </div>

We introduce a novel policy composition approach, Modality-Composable Diffusion Policy (MCDP), which composes distributional scores from multiple pre-trained diffusion policies (DPs) based on single visual modalities, enabling significant performance improvement without the need for additional training.

---

Note: The previously released module has been removed. MCDP is now a special case of our General Policy Composition (GPC) framework. Please refer to the GPC and documentation for the unified implementation.

👍 Citation

@article{cao2025MCDP,
  title={Modality-Composable Diffusion Policy via Inference-Time Distribution-level Composition},
  author={Cao, Jiahang and Zhang, Qiang and Guo, Hanzhong and Wang, Jiaxu and Cheng, Hao and Xu, Renjing},
  journal={arXiv preprint arXiv:2503.12466},
  year={2025}
}

🏷️ License

This repository is released under the MIT license. See LICENSE for additional details.