DVava/assettoCorsaGym

---
annotations_creators:
- machine-generated
language:
- en
language_creators:
- expert-generated
license:
- cc-by-4.0
multilinguality: []
pretty_name: AssettoCorsaGym Dataset
size_categories:
- 10M<n<100M
source_datasets:
- original
tags:
- RL
- MBRL
- autonomous driving
- racing
- MPC
task_categories:
- other
task_ids: []
---
# Dataset Card for Assetto Corsa Gym

## Table of Contents
- [Table of Contents](#table-of-contents)
- [Dataset Description](#dataset-description)
- [Dataset Summary](#dataset-summary)
- [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
- [Languages](#languages)
- [Dataset Structure](#dataset-structure)
- [Data Instances](#data-instances)
- [Data Fields](#data-fields)
- [Data Splits](#data-splits)
- [Dataset Creation](#dataset-creation)
- [Curation Rationale](#curation-rationale)
- [Source Data](#source-data)
- [Annotations](#annotations)
- [Personal and Sensitive Information](#personal-and-sensitive-information)
- [Considerations for Using the Data](#considerations-for-using-the-data)
- [Social Impact of Dataset](#social-impact-of-dataset)
- [Discussion of Biases](#discussion-of-biases)
- [Other Known Limitations](#other-known-limitations)
- [Additional Information](#additional-information)
- [Dataset Curators](#dataset-curators)
- [Licensing Information](#licensing-information)
- [Citation Information](#citation-information)
- [Contributions](#contributions)

## Dataset Description

- **Homepage:** https://dasgringuen.github.io/assetto_corsa_gym/
- **Repository:** https://github.com/dasGringuen/assetto_corsa_gym
- **Paper:**
- **Leaderboard:**
- **Point of Contact:** [e-mail]

### Dataset Summary

The AssettoCorsaGym dataset comprises 64 million steps, including 2.3 million steps from human drivers and the remaining from Soft Actor-Critic (SAC) policies. Data collection involved 15 drivers completing at least five laps per track and car. Participants included a professional e-sports driver, four experts, five casual drivers, and five beginners.

### Supported Tasks and Leaderboards

- Autonomous driving
- Reinforcement learning
- Behavior cloning
- Imitation learning

### Languages

English

## Dataset Structure

See https://github.com/dasGringuen/assetto_corsa_gym/blob/main/data/paths.yml and https://github.com/dasGringuen/assetto_corsa_gym/blob/main/data/README.md

```
<track>
<car>
<human / policy>
laps
```

### Data Instances

Each data instance includes telemetry data at 50Hz from a racing simulator, such as speed, position, acceleration, and control inputs (steering, throttle, brake).

### Data Fields
See:
https://github.com/dasGringuen/assetto_corsa_gym/blob/main/assetto_corsa_gym/assetto-corsa-autonomous-racing-plugin/plugins/sensors_par/structures.py

### Data Splits

We split the data in cars and tracks

## Dataset Creation

### Curation Rationale
The Assetto Corsa Gym dataset was curated to advance research in autonomous driving, reinforcement learning, and imitation learning. By providing a diverse dataset that includes both human driving data and data generated by Soft Actor-Critic (SAC) policies

### Source Data

#### Initial Data Collection and Normalization

Data was collected from a racing simulator set up. Human drivers completed at least five laps per track and car, while SAC policies were trained from scratch and their replay buffers were recorded.

#### Who are the source language producers?

Human drivers of varying skill levels, including a professional e-sports driver, experts, casual drivers, and beginners.

### Annotations

#### Annotation process

Data was automatically labeled during collection to differentiate between human and SAC policy data.

#### Who are the annotators?

The data was annotated by the research team at UC San Diego and Graz University of Technology.

### Personal and Sensitive Information

The dataset does not contain any personally identifiable information. Drivers were anonymized and identified only by driver_id.

## Considerations for Using the Data

### Social Impact of Dataset

The dataset aims to contribute to the development of safer and more efficient autonomous driving systems by providing diverse driving data for training machine learning models.

### Discussion of Biases

The dataset includes a wide range of driving skills, but there may still be biases based on the limited number of human participants and their specific driving styles. Additionally, the number of laps per track and car is unbalanced, which might affect the generalizability of models trained on this dataset. The selection of tracks and cars, as well as the specific conditions under which the data was collected, could also introduce biases that researchers should be aware of when using this dataset.

### Other Known Limitations

- Limited number of tracks and cars
- Simulated driving environment may not fully capture real-world driving conditions

## Additional Information

### Dataset Curators

The dataset was curated by researchers at UC San Diego and Graz University of Technology.

### Licensing Information

CC BY 4.0

<!-- ### Citation Information -->

### Contributions

Thanks to [@dasGringuen](https://github.com/dasGringuen) for adding this dataset.