This repository is a PyTorch implementation for paper MATT--Diff: Multimodal Active Target Tracking by Diffusion Policy Authors: Saida Liu, Nikolay Atanasov, Shumon Koga. If you are using the code for research work, please cite:

@misc{liu2025mattdiffmultimodalactivetarget,
      title={MATT-Diff: Multimodal Active Target Tracking by Diffusion Policy}, 
      author={Saida Liu and Nikolay Atanasov and Shumon Koga},
      year={2025},
      eprint={2511.11931},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2511.11931}, 
}

MATT-Diff is a diffusion-policy-based framework for active multi-target tracking in partially observable indoor environments. It learns multi-modal action strategies from three expert planners, using a preformer map encoder and attention-based target encoder over Gaussian beliefs.

Diffusion policy, Active Target Tracking, Reinforcement Learning

This repository is designed to run inside a VS Code Dev Container.

We provide:

• .devcontainer/devcontainer.json — VS Code Remote Container configuration
• Dockerfile — CUDA + Python + PyTorch + dependencies

• Docker (with NVIDIA Container Toolkit)
• NVIDIA GPU with CUDA 12 support (tested with CUDA 12.8 runtime)
• VS Code + "Dev Containers" / "Remote - Containers" extension

1. Clone the repository:

   git clone https://github.com/CINAPSLab/MATT-Diff.git
   cd MATT-Diff

2. Open this folder in VS Code and click:

   “Reopen in Container”

3. The dev container will automatically:

   • Build from the provided Dockerfile (CUDA 12.8 / Ubuntu 22.04)
   • Create a virtual environment at /opt/venv
   • Install nightly PyTorch with CUDA 12.8 + all dependencies from requirements.txt
   • Set environment variables such as PYTHONPATH=/workspace:/workspace/src

We provide pretrained checkpoints via HuggingFace. Training from scratch requires the expert dataset, which is not included in this repository.

huggingface-cli download saidaliu27/MATT-Diff dp/best.pt --local-dir src/run/dp
huggingface-cli download saidaliu27/MATT-Diff bc/best.pt --local-dir src/run/bc

curl -L https://huggingface.co/saidaliu27/MATT-Diff/resolve/main/dp/best.pt -o src/run/dp/best.pt
curl -L https://huggingface.co/saidaliu27/MATT-Diff/resolve/main/bc/best.pt -o src/run/bc/best.pt

After downloading, you can directly evaluate MATT-Diff

python -u -m evals.eval_dp

python -u -m logger.explore_logger # collect data of Frontier-based planner
python -u -m logger.mm_logger # collect data of Time-based, Uncertainty-based planner
python -u -m logger.refinedata # refine collected expert data
python -u -m src.train_dp # train MATT-diff