$\textsf{\color{orange}{M}\color{orange}{P}\textcolor{orange}{1}}$: $\textsf{\color{orange}{M}}\text{ean}$ Flow Tames $\textsf{\color{orange}{P}}\text{olicy}$ Learning in $\textsf{\color{orange}{1}}$-step for Robotic Manipulation

Juyi Sheng^*, Ziyi Wang^*, Peiming Li, Mengyuan Liu^†

State Key Laboratory of General Artificial Intelligence, Peking University, Shenzhen Graduate School, Shenzhen, China

Project Page | arXiv

Abstract

In robot manipulation, robot learning is becoming a prevailing approach. However, generative models within this field face a fundamental trade-off between the slow, iterative sampling of diffusion models and the architectural constraints of faster Flow-based methods, which often rely on explicit consistency losses. To address these limitations, we introduce MP1, which pairs 3D point-cloud inputs with the MeanFlow paradigm to generate action trajectories in one network function evaluation (1-NFE). By directly learning the interval-averaged velocity via the “MeanFlow Identity,” our policy avoids any additional consistency constraints. This formulation eliminates numerical ODE-solver errors during inference, yielding more precise trajectories. MP1 further incorporates CFG for improved trajectory controllability while retaining 1-NFE inference without reintroducing structural constraints. Because subtle scene-context variations are critical for robot learning, especially in few-shot learning, we introduce a lightweight Dispersive Loss that repels state embeddings during training, boosting generalization without slowing inference. We validate our method on the Adroit and Meta-World benchmarks, as well as in real-world scenarios. Experimental results show MP1 achieves superior average task success rates, outperforming DP3 by 10.2% and FlowPolicy by 7.3%. Its average inference time is only 6.8 ms—19× faster than DP3 and nearly 2× faster than FlowPolicy. Our code is available at https://mp1-2254.github.io/.

Pipeline

Overview of MP1. The MP1 takes the historical observation point cloud and the robot's state as inputs. These inputs are processed through a visual encoder and a state encoder, respectively, and then serve as conditional inputs to the UNet-integrated MeanFlow. After passing through the MeanFlow, the model computes regression loss ($L_{cfg}$) between the mean velocity generated from the initial noise and the target velocity. This $L_{cfg}$ is combined with a Dispersive Loss ($L_{disp}$) imposed on the UNet’s hidden states to jointly optimize the network parameters.

💻 Installation

See install.md for installation instructions.

📚 Data

You could generate demonstrations by yourself using our provided expert policies. Generated demonstrations are under $YOUR_REPO_PATH/MP1/data/.

🛠️ Usage

Scripts for generating demonstrations, training, and evaluation are all provided in the scripts/ folder.

The results are logged by swanlab or wandb, so you need to swanlab login first to see the results.

1. Generate demonstrations by gen_demonstration_adroit.sh and gen_demonstration_metaworld.sh. See the scripts for details. For example:

   bash scripts/gen_demonstration_metaworld.sh drawer-close

   This will generate demonstrations for the drawer-close task in Meta-World environment. The data will be saved in MP1/data/ folder automatically.

2. Train and evaluate a policy with behavior cloning. For example:

   bash ./auto_run.bash 0 mp

   This will train an MP/MP1 policy on the drawer-close task in Meta-World environment using point cloud modality.

🤖 Real-world Deploy

R-MP1

🏷️ License

This repository is released under the MIT license.

🙏 Acknowledgement

Our code is built upon 3D Diffusion Policy, Flow Policy, Mean Flow, Mean Flow official, VRL3, and Metaworld. We would like to thank the authors for their excellent works.

👍 Citation

@article{sheng2025mp1meanflowtames,
      title={MP1: Mean Flow Tames Policy Learning in 1-step for Robotic Manipulation}, 
      author={Juyi Sheng and Ziyi Wang and Peiming Li and Mengyuan Liu},
      year={2025},
      eprint={2507.10543},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2507.10543}, 
}