Guowei-Zou/dm1-pretrained

DM1: MeanFlow with Dispersive Regularization for 1-Step Robotic Manipulation

Model Description

DM1 is a novel flow matching framework for robotic manipulation that achieves one-step inference while maintaining high success rates. The model prevents representation collapse through dispersive regularization while achieving 20-40× faster inference compared to diffusion baselines.

Key Features

• ⚡ Single-Step Inference: 0.07s per timestep (vs. 2-3.5s for diffusion)
• 🎯 High Success Rates: 10-20% improvement over diffusion baselines
• 🔧 Dispersive Loss Family: InfoNCE, Cosine, Hinge regularizers
• 👁️ Vision-Ready: Multi-view RGB observation support
• 🤖 Real Robot Validated: Tested on Franka-Emika-Panda

Model Variants

This repository contains pretrained weights for multiple configurations:

Variant	Description	Best For
ShortCut + InfoNCE L2	Flow matching with L2-based dispersive loss	General tasks
ShortCut + InfoNCE Cosine	Flow matching with cosine similarity	Vision tasks
ShortCut + Hinge	Flow matching with hinge loss	Robust control
ShortCut + Covariance	Flow matching with covariance regularization	Feature diversity
MeanFlow variants	Mean flow baseline and dispersive versions	Fast inference
ReFlow variants	Reflow baseline	Iterative refinement

Supported Tasks

• Robomimic (RGB): lift, can, square, transport
• Franka Kitchen: partial, complete, mixed
• D3IL: avoiding, pushing, sorting

Quick Start

Installation

git clone https://github.com/Guowei-Zou/dm1-release.git
cd dm1-release
conda create -n dm1 python=3.8 -y
conda activate dm1
pip install -e .

Download Checkpoints

from huggingface_hub import hf_hub_download

# Download specific checkpoint
checkpoint = hf_hub_download(
    repo_id="zougw2025/dm1-pretrained",
    filename="checkpoints/w_0p1/can/can_w0p1_05_shortcut_infonce_cosine.pt"
)

Evaluation

python script/run.py \
  --config-dir=cfg/robomimic/eval/can \
  --config-name=eval_shortcut_mlp_img \
  base_policy_path=checkpoints/w_0p1/can/can_w0p1_05_shortcut_infonce_cosine.pt

Performance Metrics

Task	Baseline (32-128 steps)	DM1 (5 steps)	Improvement	Speedup
Lift	~85%	99%	+14%	20-40×
Can	Variable	High	+10-20%	20-40×
Square	Moderate	Improved	+15-25%	20-40×
Transport	Low	High	+20-30%	20-40×

Citation

If you use DM1 in your research, please cite:

@misc{zou2025dm1meanflowdispersiveregularization,
      title={DM1: MeanFlow with Dispersive Regularization for 1-Step Robotic Manipulation},
      author={Guowei Zou and Haitao Wang and Hejun Wu and Yukun Qian and Yuhang Wang and Weibing Li},
      year={2025},
      eprint={2510.07865},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2510.07865},
}

License

This project is licensed under the Apache License 2.0.

Acknowledgments

DM1 builds upon prior work including Diffusion Policy, ReinFlow, MeanFlow, FlowPolicy, D2PPO, and π0.5.

Contact

• Author: Guowei Zou
• GitHub: dm1-release
• Project Page: https://guowei-zou.github.io/dm1/