Unified Generation-Refinement Planning: Bridging Guided Flow Matching and Sampling-Based MPC for Social Navigation

Kazuki Mizuta¹ · Karen Leung^1,2

¹University of Washington     ²NVIDIA

---

Installation

To set up the environment, run the following commands:

conda env create -f conda_environment.yaml
conda activate cfm_mppi
pip install -e . 

Usage

Training

To train the CFM (Conditional Flow Matching) model, run:

python cfm_mppi/train.py

Demos

We provide Jupyter notebook demos for different dynamics models:

• Unicycle dynamics: cfm_mppi/example/unicycle.ipynb
• Double integrator dynamics: cfm_mppi/example/doubleintegrator.ipynb

Pre-trained Models and Datasets

For quick evaluation, download the pre-trained weights and datasets from Google Drive and organize them into the following directory structure:

• dataset/ → ./dataset/
• checkpoint.pt → ./output_dir/cfm_transformer/checkpoint.pth
• args.json → ./output_dir/cfm_transformer/args.json

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{MizutaLeung2026,
  title = {Unified Generation-Refinement Planning: Bridging Guided Flow Matching and Sampling-Based MPC for Social Navigation},
  author = {Kazuki Mizuta and Karen Leung},
  booktitle = {Proc.\ IEEE Conf.\ on Robotics and Automation},
  year = {2026},
}

Acknowledgements

This repository is built upon several excellent open-source projects:

• flow_matching
• latent-diffusion
• diffuser
• trajdata
• mppi_playground