Dynamic Objects Relocalization in Changing Environments with Flow Matching

Francesco Argenziano^1,# · Miguel Saavedra-Ruiz^2,3,# · Sacha Morin^2,3 · Daniele Nardi¹ · Liam Paull^2,3
1Department of Computer, Control and Management Engineering, Sapienza University of Rome, Italy
²DIRO, Université de Montréal, Montréal, QC, Canada · ³Mila - Quebec AI Institute, Montréal, QC, Canada
^#Co‑first authors

---

Official code release for the paper Dynamic Objects Relocalization in Changing Environments with Flow Matching (code‑name: FlowMaps).

Table of Contents

• Installation
• Dataset Generation (FlowSim)
• Configuration
• Training & Logging
• Evaluation & Reproducing Results
• Citation
• License

Installation

# 1) Clone
git clone https://github.com/Fra-Tsuna/flowmaps
cd flowmaps

# 2) Create an environment (example with conda) and install deps
conda create -n flowmaps python=3.13 -y
conda activate flowmaps
pip install -r requirements.txt

Dataset Generation (FlowSim)

To generate datasets with FlowSim:

python3 data.py

This creates training/validation data according to config/config.yaml. We use a Hydra‑style syntax for overrides (key=value). You can edit config/config.yaml directly or pass overrides at the CLI.

python3 data.py n_env_train=200 n_env_val=20 max_timesteps=20 stochastic=true

Key parameters:

• n_env_train, n_env_val: number of training and validation environments (respectively).
• max_timesteps: time horizon of object trajectories.
• stochastic: enable stochastic object behavior.
• size: environment canvas size.
• max_tables, min_each: number of furniture instances and minimum per type.
• display_scale: visualization only; upscales the rendered image. Underlying data remain at size×size.

Training & Logging

Our project supports Weights & Biases logging.

• Online (default below):

  python3 main.py wandb.mode=online wandb.project=YOUR_PROJECT wandb.entity=YOUR_ENTITY wandb.tags="[flowmaps,cdit]" experiment=cdit

• Offline:

  python3 main.py wandb.mode=offline experiment=cdit

To train the MLP baseline, simply override experiment=mlp

Evaluation & Reproducing Results

After training, place your checkpoint file in ckpt/ and run:

python3 eval.py checkpoint_name=YOUR_CHECKPOINT_FILENAME

Citation

If you find this work useful, please cite the paper:

@article{argenziano2025dynamic,
  title={Dynamic Objects Relocalization in Changing Environments with Flow Matching},
  author={Argenziano, Francesco and Saavedra-Ruiz, Miguel and Morin, Sacha and Nardi, Daniele and Paull, Liam},
  journal={arXiv preprint arXiv:2509.16398},
  year={2025}
}

License

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