README.md

---
pretty_name: TacSIm
task_categories:
- reinforcement-learning
tags:
- multi-agent
- imitation-learning
- football
- soccer
- trajectory
- tactical-imitation
- sequential-data
- google-research-football
- benchmark
configs:
- config_name: default
  data_files:
  - split: train
    path: Train/**/*.csv
  - split: validation
    path: Validation/**/*.csv
  - split: test
    path: Test/**/*.csv
---

# TacSIm: A Dataset and Benchmark for Football Tactical Style Imitation

## Dataset Description

**TacSIm** is a multi-agent football trajectory dataset designed for research on tactical style imitation, multi-agent imitation learning, trajectory generation, and long-horizon football simulation.

The dataset is organized into three predefined splits:

- **Train**
- **Validation**
- **Test**

These splits should be preserved when training and evaluating models so that reported results remain comparable.

> **Paper:** *TacSIm: A Dataset and Benchmark for Football Tactical Style Imitation*  
> **Venue:** CVPR 2026  
---

## Dataset Structure

### Recommended Repository Layout

```text
TacSIm/
├── README.md
├── Train/
│   ├── 00001.csv
│   ├── 00002.csv
│   └── ...
├── Validation/
│   ├── 00001.csv
│   ├── 00002.csv
│   └── ...
└── Test/
    ├── 00001.csv
    ├── 00002.csv
    └── ...
```

Each CSV file represents one football trajectory segment.

### Trajectory Length

In the released format:

- each file contains **100 time steps**;
- the sampling interval is **0.1 seconds**;
- each file therefore represents a **10-second trajectory**;
- rows are ordered chronologically.

A model may be evaluated at shorter horizons by using the first:

| Evaluation horizon | Number of rows |
|---|---:|
| 3 seconds | 30 |
| 5 seconds | 50 |
| 10 seconds | 100 |

---


### Player Positions

For each controlled player \(i \in \{1,\ldots,11\}\):

```text
player_i_x
player_i_y
```

These columns contain the player's normalized two-dimensional position at each time step.

`player_1` is treated as the goalkeeper in the TacSIm team-level evaluation protocol. Players `player_2` through `player_11` are treated as outfield players.


### Complete Column List

```text
player_1_x, player_1_y,
player_2_x, player_2_y,
player_3_x, player_3_y,
player_4_x, player_4_y,
player_5_x, player_5_y,
player_6_x, player_6_y,
player_7_x, player_7_y,
player_8_x, player_8_y,
player_9_x, player_9_y,
player_10_x, player_10_y,
player_11_x, player_11_y,
ball_x, ball_y,
player_1_action,
player_2_action,
player_3_action,
player_4_action,
player_5_action,
player_6_action,
player_7_action,
player_8_action,
player_9_action,
player_10_action,
player_11_action
```

---

## Coordinate Convention

The coordinates follow the normalized Google Research Football field representation.

The nominal coordinate ranges are approximately:

```text
x: [-1.00, 1.00]
y: [-0.42, 0.42]
```

Small values outside these nominal limits may occur because of simulator dynamics, boundary behavior, or preprocessing. Users should avoid silently clipping coordinates unless clipping is explicitly part of their experimental protocol.

The released files contain only the two-dimensional coordinates used by the TacSIm benchmark. Ball height is not included in this CSV format.

---

## Loading the Dataset

### Load with Hugging Face Datasets

```python
from datasets import load_dataset

dataset = load_dataset("YOUR_USERNAME/TacSIm")

print(dataset)
print(dataset["train"][0])
```

The YAML configuration at the top of this card maps the repository folders to the following Hugging Face splits:

```text
Train      -> train
Validation -> validation
Test       -> test
```

### Important Note About Trajectory Boundaries

When multiple CSV files are loaded through `datasets.load_dataset`, Hugging Face may concatenate their rows within each split. The original source filename is not necessarily preserved as a feature.

For methods that require each 100-frame trajectory to remain a separate episode, download the repository while preserving its directory structure:

```python
from pathlib import Path

import pandas as pd
from huggingface_hub import snapshot_download

dataset_root = Path(
    snapshot_download(
        repo_id="YOUR_USERNAME/TacSIm",
        repo_type="dataset",
    )
)

train_files = sorted((dataset_root / "Train").glob("*.csv"))
validation_files = sorted(
    (dataset_root / "Validation").glob("*.csv")
)
test_files = sorted((dataset_root / "Test").glob("*.csv"))

trajectory = pd.read_csv(train_files[0])

print(train_files[0].name)
print(trajectory.shape)
```

### Load One Trajectory

```python
import pandas as pd

trajectory = pd.read_csv("Train/02304.csv")

player_positions = trajectory[
    [
        column
        for column in trajectory.columns
        if column.startswith("player_")
        and (
            column.endswith("_x")
            or column.endswith("_y")
        )
    ]
]

ball_positions = trajectory[["ball_x", "ball_y"]]

player_actions = trajectory[
    [
        column
        for column in trajectory.columns
        if column.endswith("_action")
    ]
]
```

### Convert One File to Arrays

```python
import numpy as np
import pandas as pd

trajectory = pd.read_csv("Train/02304.csv")

player_columns = []

for player_id in range(1, 12):
    player_columns.extend(
        [
            f"player_{player_id}_x",
            f"player_{player_id}_y",
        ]
    )

action_columns = [
    f"player_{player_id}_action"
    for player_id in range(1, 12)
]

players = trajectory[player_columns].to_numpy(
    dtype=np.float32
).reshape(-1, 11, 2)

ball = trajectory[["ball_x", "ball_y"]].to_numpy(
    dtype=np.float32
)

actions = trajectory[action_columns].to_numpy(
    dtype=np.int64
)

print("Players:", players.shape)  # (100, 11, 2)
print("Ball:", ball.shape)        # (100, 2)
print("Actions:", actions.shape)  # (100, 11)
```

---

## Dataset Splits

The dataset contains fixed training, validation, and test partitions.

| Split | Purpose | Number of trajectory files |
|---|---|---:|
| Train | Model training | 10756 |
| Validation | Hyperparameter selection and model development | 2304 |
| Test | Final benchmark evaluation | 2304 |

The test set should not be used for model selection or hyperparameter tuning.

To count the files locally:

```python
from pathlib import Path

for split in ["Train", "Validation", "Test"]:
    count = len(list(Path(split).glob("*.csv")))
    print(split, count)
```

---

## Benchmark Tasks

TacSIm supports research on:

1. **Multi-agent behavior cloning**
2. **Multi-agent imitation learning**
3. **Football trajectory prediction**
4. **Tactical style imitation**
5. **Long-horizon multi-agent simulation**
6. **World-model learning**
7. **Offline reinforcement learning**
8. **Multi-agent action prediction**

Depending on the experimental setting, a model may use historical player positions, ball positions, and player actions to predict future trajectories or future joint actions.

---

## Benchmark Evaluation

The TacSIm benchmark evaluates generated trajectories at:

```text
3 seconds
5 seconds
10 seconds
```

Ball-trajectory similarity is evaluated under multiple spatial grid resolutions:

```text
10 × 6
15 × 10
20 × 12
30 × 20
105 × 68
```

The benchmark includes the following trajectory-level measurements:

- **Spatial Occupancy Similarity**
- **Movement Vector Similarity**

Extended team-level evaluation may additionally report:

- **Team Formation Compactness Similarity**
- **Team Speed Similarity**

For reproducible comparisons, all methods should use the same:

- train/validation/test split;
- temporal horizons;
- coordinate convention;
- grid resolutions;
- goalkeeper handling;
- aggregation procedure.

---

## Intended Uses

The dataset is intended for academic research in:

- multi-agent learning;
- sports analytics;
- football simulation;
- imitation learning;
- trajectory modeling;
- tactical behavior generation;
- world models;
- sequential decision making.

The dataset may also be used for teaching and reproducibility studies related to multi-agent football environments.

---


## Data Quality and Preprocessing

Each trajectory should be checked for:

- exactly 100 ordered time steps;
- all 35 required columns;
- finite coordinate values;
- action IDs within 0–18;
- consistent coordinate orientation;
- consistent player indexing.

A basic validation example is:

```python
from pathlib import Path

import numpy as np
import pandas as pd

required_columns = [
    item
    for player_id in range(1, 12)
    for item in (
        f"player_{player_id}_x",
        f"player_{player_id}_y",
    )
] + [
    "ball_x",
    "ball_y",
] + [
    f"player_{player_id}_action"
    for player_id in range(1, 12)
]

for csv_file in Path("Train").glob("*.csv"):
    data = pd.read_csv(csv_file)

    assert len(data) == 100
    assert all(
        column in data.columns
        for column in required_columns
    )
    assert np.isfinite(
        data[required_columns].to_numpy()
    ).all()

    actions = data[
        [
            f"player_{player_id}_action"
            for player_id in range(1, 12)
        ]
    ].to_numpy()

    assert actions.min() >= 0
    assert actions.max() <= 18
```

---

## Ethical Considerations

TacSIm contains simulated trajectories and does not contain personal information, biometric identifiers, or recordings of real individuals.

Researchers should clearly distinguish simulated findings from conclusions about real athletes or teams.

---

## Citation

Please cite the TacSIm paper when using this dataset:

```bibtex
@inproceedings{wen2026tacsim,
  title={TacSIm: A Dataset and Benchmark for Football Tactical Style Imitation},
  author={Wen, Peng and Wang, Yuting and Wang, Qiurui},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={20014--20023},
  year={2026}
}
```

---

## Contact

For questions, issues, or benchmark submissions, please contact:

```text
Name:Peng Wen
Email: wenpengsc@gmail.com

```