README.md

---
license: cc-by-nc-sa-4.0
pretty_name: Rocket Science
language:
  - en
size_categories:
  - 1M<n<10M
task_categories:
  - other
tags:
  - rocket-league
  - world-model
  - video
  - reinforcement-learning
  - multimodal
  - game
  - webdataset
extra_gated_prompt: >-
  Rocket League © Psyonix LLC / Epic Games, Inc. Rocket League and its logo are trademarks of
  Psyonix LLC. The Rocket League gameplay recordings in this dataset are used and distributed with
  Epic Games' permission. This dataset is licensed under CC BY-NC-SA 4.0. Per Section 2(b)(2) of that
  license, no trademark, publicity, or other rights are granted beyond the copyright license to this
  dataset. Commercial use of Rocket League content requires separate permission from Epic Games. This
  dataset and the associated research are not an Epic Games product. All gameplay was generated by
  bots in private matches; the dataset contains no human player data. You must accept these terms to
  access the dataset.
extra_gated_fields:
  I agree to use this dataset for non-commercial purposes only, under CC BY-NC-SA: checkbox
  I acknowledge Rocket League content is © Psyonix/Epic and that commercial use requires separate permission from Epic Games: checkbox
---

# Rocket Science

Time-aligned **video, keyboard actions, game events, and per-frame game state** for all four players, captured from a 2v2 Rocket League match.

This is the dataset behind **MIRA**, a real-time multiplayer world model trained to simulate Rocket League gameplay — by [General Intuition](https://www.generalintuition.com) and [Kyutai](https://kyutai.org), in collaboration with [Epic Games](https://www.epicgames.com). Code, technical report,
and a live demo: [github.com/mira-wm/mira](https://github.com/mira-wm/mira) ·
[mira-wm.com](https://mira-wm.com).

## Loading

```python
from mira.data import RocketScienceDataset

ds = RocketScienceDataset.from_hub("kyutai/rocket-science", split="test")
clip = ds.load_match(ds.match_ids()[0], clip_len=16, target_fps=10)[0]
```

Install the loader with `pip install "mira[hf,decode] @ git+https://github.com/mira-wm/mira"`
(provide your own torch + a compatible FFmpeg), or use the repo's pixi environment.

## What's in a sample

One sample is a short (~4 s, 80 frames at 20 fps) window of a match with all four players'
synchronised views. A **clip** is a fixed-length, fps-downsampled slice read from one. For each clip,
per perspective:

- **`frames`** — `(P, T, C, H, W)` uint8 video (1280×720), one decoded view per perspective.
- **`actions`** — `(P, T, 9)` int32 multi-hot keyboard over a fixed key set
  (`W A S D Q E Space LShiftKey LControlKey`), OR-ed over each downsample window so it stays
  frame-aligned.
- **`events`** — discrete game events (`KickoffStarted/Ended`, `GoalScored`, `Demolition`,
  `GoalReplayStarted/Ended`, …) on a single match-wide master clock, mapped onto each perspective's
  frames via its `recording_offset_sec`. `exclude_replays=True` skips clips overlapping goal replays.
- **`physics`** — per-frame game state, one list of `FrameState` dicts per perspective, frame-aligned
  with the video: the **ball**, the four **cars**, and the **game** info (score, overtime, clock).
  See `mira.data.state` for the typed field layout.

Perspectives are ordered by `player_id`; `clip.teams` gives each one's team (0 / 1). The video is
H.264, 1280×720 at 20 fps, no audio.

## Layout

```
kyutai/rocket-science
├── test/
│   ├── index.json          # per-match index: shard path, perspectives, chunk frame counts
│   └── dataset_*.tar       # WebDataset shards
└── train/
    ├── index.json
    └── 000/ … 015/         # shards in numbered subfolders of ≤1,000 files
        └── dataset_*.tar
```

| Split | Matches | Chunks    | Shards | Size    |
|-------|---------|-----------|--------|---------|
| train | 15,769  | 1,792,926 | 15,274 | ~27 TB  |
| test  | 62      | 7,187     | 61     | ~109 GB |

Each WebDataset sample is one ~4 s chunk of one match (key `{match_id}_c{chunk:05d}`), bundling all
four perspectives: `p{i}.mp4` (video), `p{i}.jsonl` (actions), `p{i}.physics.jsonl` (game state,
one JSON line per frame), and `meta.json` (`match_id`, `chunk`, `player_ids`, `teams`). The loader
resolves shards through `index.json`, so the train subfolders are transparent; any WebDataset-aware
pipeline works too, e.g.
`datasets.load_dataset("webdataset", data_files="hf://datasets/kyutai/rocket-science/train/*/*.tar")`
(for test: `.../test/*.tar`).

## License & attribution

Rocket League © Psyonix LLC / Epic Games, Inc. Rocket League and its logo are trademarks of Psyonix
LLC. The Rocket League gameplay recordings in this dataset are used and distributed with Epic Games'
permission.

Licensed under **CC BY-NC-SA 4.0**. Per Section 2(b)(2), no trademark, publicity, or other rights are
granted beyond the copyright license to this dataset; commercial use of Rocket League content
requires separate permission from Epic Games. This dataset and the associated research are not an
Epic Games product. All gameplay was generated by bots in private matches; the dataset contains no
human player data.

## Citation

If you find this dataset useful, please cite our [technical report](https://arxiv.org/abs/2607.05352):

```bibtex
@misc{hu2026multiplayerinteractiveworldmodels,
      title={Multiplayer Interactive World Models with Representation Autoencoders},
      author={Anthony Hu and Václav Volhejn and Adrien Ramanana Rahary and Chris Mulder and Aditya Makkar and Alyx Liao and Amélie Royer and Manu Orsini and Adam Jelley and Eloi Alonso and Florian Laurent and Fredrik Norén and James Swingos and Jan Hünermann and Kent Rollins and Lucas Hosseini and Matthieu Le Cauchois and Maxim Peter and Pim de Witte and Tim Brown and Vincent Micheli and Moritz Böhle and Gabriel de Marmiesse and Viktoriia Sharmanska and Lucia Specia and Michael Black and Patrick Pérez},
      year={2026},
      eprint={2607.05352},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.05352},
}
```