README.md

---
pretty_name: FMUKF Containership Dataset
license: cc-by-4.0
tags:
- simulation
- time-series
- maritime
- control
- state-estimation
- ukf
- foundation-models
- hdf5
- fmukf
task_categories:
- time-series-forecasting
size_categories:
- 100K<n<1M
---

# FMUKF Containership Dataset

## Dataset Summary

**Containerships-Large** is the dataset used in our paper:  
> *Sailing Towards Zero-Shot State Estimation using Foundation Models Combined with a UKF* (CDC 2025, preprint available on [Arxiv](https://arxiv.org/abs/2509.04213) ).  

Please refer to the papers's [GitHub](https://github.com/Data-Science-in-Mechanical-Engineering/fm-ukf) repository and [Arxiv](https://arxiv.org/abs/2509.04213) for full code and details.

It contains **1,000** unique parameterizations of a 6-DOF containership model and **400** simulated trajectories per parameterization (each 384 seconds long), useful for training and testing foundation models.  
The base ship model follows [Son & Nomoto (1981](https://www.jstage.jst.go.jp/article/jjasnaoe1968/1981/150/1981_150_232/_article) and the implementation are adapted from on parameters/dynamics from Thor I. Fossen’s [Marine System Simulator (MSS)](https://github.com/cybergalactic/MSS/blob/master/VESSELS/models/shipModels/dataContainer.m). 

- **File format:** single HDF5 file `Containerships_Large.h5`  
- **State vector X shape:** `(384, 10)`  
- **Control vector U shape:** `(384, 2)`

## Supported Tasks & Benchmarks

- **State estimation / filtering**
- **System identification & dynamics modeling**  
- **Time-series forecasting / foundation models**

## Dataset Structure

- **Total environments (ships):** 1,000  
- **Trajectories per environment:** 400  
- **Timesteps per trajectory:** 384 (1 s per step)

```
Containerships_Large.h5
├── env_0/                     # First Ship Parametrization
│   ├── .attrs["parameters"]   # YAML string encoding parameters
│   ├── traj_0/                
│   │   ├── x                  # (384, 10) state trajectory
│   │   └── u                  # (384, 2) control inputs
│   ├── traj_1/
│   │   ├── x
│   │   └── u
│   └── ...
├── env_1/
│   ├── .attrs["parameters"]
│   ├── traj_0/  ...
└── ...
```

### Data Fields

**State `x` (10-dim):**
- `u` — surge velocity [m/s]  
- `v` — sway velocity [m/s]  
- `r` — yaw rate [deg/s]  
- `x` — North position [km]  
- `y` — East position [km]  
- `psi` — yaw angle [deg]  
- `p` — roll rate [deg/s]  
- `phi` — roll angle [deg]  
- `delta` — rudder angle [deg]  
- `n` — propeller shaft velocity [rpm]

**Control `u` (2-dim,):**
- `delta` — commanded rudder angle [deg]  
- `n` — commanded propeller shaft velocity [rpm]

### Parameter Variations

We sample **1,000** ship parameterizations by perturbing key physical parameters by ≈±30% around the original base model. We filter unstable or near-duplicate systems using a similarity metric to ensure diversity and stability. Each trajectory starts from a randomized initial state; control inputs are pink noise.

## How to Use

### Quick start (download via `huggingface_hub`)
```python
from huggingface_hub import hf_hub_download
import h5py, yaml, numpy as np

# replace with your dataset repo id
repo_id = "your-username/containerships-large"
path = hf_hub_download(repo_id=repo_id, filename="Containerships_Large.h5", repo_type="dataset")

with h5py.File(path, "r") as f:
    env_keys = list(f.keys())

    # Select env (ship parametrization) and trajectory number
    env_key = env_keys[0]  #<-- "env_0", "env_1", ..., "env_999"
    traj_key = "traj_0"    #<-- "traj_0", "traj_1", ..., "traj_399"

    # Get State Vector and Control Inputs
    X = np.array(f[env_key][traj_key]["x"])  # (384, 10)
    U = np.array(f[env_key][traj_key]["u"])  # (384, 2)

    # Load parameters
    params = yaml.safe_load(f[env_key].attrs["parameters"])

```

## Data Generation & Provenance

## Licensing

**Dataset license:** **CC-BY-4.0** — you may use, share, and adapt the data with attribution to the authors.  

## Intended Uses & Limitations

- **Intended uses:** research on filtering, state estimation, dynamics learning, forecasting, and control.  
- **Not for safety-critical use:** do **not** use to design or validate real-world maritime operations without expert verification and domain-specific validation.

## Maintenance

- **Authors:** Tobin Holtmann, Data Science in Mechenical Engineering, RWTH Aachen University
- **Version:** v1.0 (2025-09-03)  

## Citation

Please cite our paper and acknowledge the dataset:

@article{holtmann2025fmukf,
  title     = {Sailing Towards Zero-Shot State Estimation using Foundation Models Combined with a UKF},
  author    = {Holtmann, Tobin and Stenger, David and Posada-Moreno, Andres and Solowjow, Friedrich and Trimpe, Sebastian},
  journal   = {arXiv preprint arXiv:2509.04213},
  year      = {2025},
  doi       = {10.48550/arXiv.2509.04213},
  note      = {Accepted for the 64th IEEE Conference on Decision and Control (CDC 2025)},
}