DnDFilter

This repository is the official implementation of the paper "DnD Filter: Differentiable State Estimation for Dynamic Systems using Diffusion Models".

DnD Filter is a differentiable filter that utilizes diffusion models for state estimation of dynamic systems. Unlike conventional differentiable filters, which often impose restrictive assumptions on process noise (e.g., Gaussianity), DnD Filter enables a nonlinear state update without such constraints by conditioning a diffusion model on both the predicted state and observational data, capitalizing on its ability to approximate complex distributions. To the best of our knowledge, DnD Filter represents the first successful attempt to leverage diffusion models for state estimation, offering a flexible and powerful framework for nonlinear estimation under noisy measurements.

Overview

This repository contains the training and validation code for DnD Filter, as well as pre-trained model checkpoints. Below is an outline of the key scripts and directories:

• ./train/train.py: Training script for DnD Filter and baseline methods.
• ./train/test_*.py: Validation scripts for DnD Filter and baseline methods.
• ./train/config/: Configuration files for training DnD Filter and baseline methods.
• ./train/dataset/: Datasets used for both training and validation.
• ./train/logs/: Directory containing trained model checkpoints for DnD Filter and baseline methods.
• ./train/DND_train/: Model implementation files for DnD Filter and its baselines.

Getting Started

Run the commands below inside the topmost directory:

1. Set up the conda environment:

   conda env create -f train_environment.yml
   

2. Source the conda environment:

   conda activate DnD_Filter
   

3. Install the diffusion_policy package from this repo into the state_estimation_* folders:

   git clone git@github.com:real-stanford/diffusion_policy.git
   pip install -e diffusion_policy/
   

Training and Validating

For training, modify the configuration file in ./train/train.py to match the desired training objective or model, then simply run train.py to start the training process.

For validating, run the test scripts found in the ./train/test_*.py or ./train/test.py files.

To train from a existing checkpoints, Add

    load_run: <project_name>/<log_run_name>

to .yaml config file in ./train/config/. The *.pth of the file you are loading to be saved in this file structure and renamed to “latest”:

   state_estimation_*/train/logs/<project_name>/<log_run_name>/latest.pth. 

Dataset

1. The simulated disk tracking dataset can be generated using codes in repo.
2. KITTI Visual Odometry Dataset (https://www.cvlibs.net/datasets/kitti/eval_odometry.php)

The dataset should be processed into following structure:

├── <dataset_name>
│   ├── <name_of_traj1>
│   │   ├── 0
│   │   ├── 1
│   │   ├── ...
│   │   ├── T-1
│   │   ├── traj_data.pkl
│   │   └── traj_data.txt
│   ├── <name_of_traj2>
│   │   ├── 0
│   │   ├── 1
│   │   ├── ...
│   │   ├── T-1
│   │   ├── traj_data.pkl
│   │   └── traj_data.txt
│   ...
└── └── <name_of_trajN>
    	├── 0
    	├── 1
    	├── ...
        ├── T-1
        ├── traj_data.pkl
        └── traj_data.txt

Files 0 to T-1 contain the high-dimensional observations (e.g.,images), while traj_data.pkl and traj_data.txt store metadata and additional information related to the sequence.

The processed dataset used in our experiments is available at DataLink. After downloading, place the dataset into the ./train/dataset directory.

Citation

@article{Wan2025DnD,
  title={DnD Filter: Differentiable State Estimation for Dynamic Systems using Diffusion Models},
  author={Ziyu Wan, Lin Zhao},
  journal={arXiv preprint arXiv:2503.01274},
  year={2025}
}