File size: 3,019 Bytes
8dcf1f9 c9a15fd 8dcf1f9 c9a15fd | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 | ---
license: cc-by-4.0
library_name: mujoco
pipeline_tag: reinforcement-learning
tags:
- reinforcement-learning
- mujoco
- humanoid
- martial-arts
- ppo
- webassembly
---
# Kakuto fighter
A simulated humanoid martial artist, trained with reinforcement learning (PPO) in [MuJoCo](https://mujoco.org).
It holds a Muay Thai stance, walks and dashes, blocks, and throws 14 techniques: jab, cross, lead and rear hooks,
uppercuts, body cross, pull counter, high kicks, knees and elbows.
**Try it in your browser:** [ai.kakuto.io/en/try/](https://ai.kakuto.io/en/try/). The page runs MuJoCo compiled to
WebAssembly and this policy entirely on your device, and downloads the files in `web/` from this repository.
How it was trained, with videos and costs: [ai.kakuto.io](https://ai.kakuto.io).
## This model
| | |
|---|---|
| Checkpoint | `distill9-196.6M` (196.6 M environment steps) |
| What it is | A student policy distilled from per-skill teacher policies |
| Body | `muscular-fast`: 48 actuators |
| Network | MLP, 2 hidden layers of 512 units, ELU; 465 observations in, 48 actions out |
## Files
Everything the browser demo needs is in `web/`:
| File | What |
|---|---|
| `fighter.xml` | The MuJoCo model (MJCF): the fighter's body and the scene |
| `fighter.json` | The metadata: tensor layout in `fighter.bin`, observation layout, action scaling, techniques |
| `fighter.bin` | The policy's weights and observation normalisation, plus the stance tables |
| `punches.bin.gz` | The techniques' reference tables (loaded when the first technique is used) |
| `walk.bin.gz` | The walking reference tables |
| `mujoco-3.16.0.wasm.gz` | MuJoCo 3.16.0's WebAssembly engine, gzipped; unchanged from the `@mujoco/mujoco` npm package |
| `MUJOCO_LICENSE` | MuJoCo's licence (Apache 2.0) |
The observation is built by the environment code on the demo page, so the weights alone do not reproduce the
behaviour.
## Licence
The fighter's model, weights and tables are © 2026 Kakuto Technology GK, under
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/): you may use, share and adapt them for any purpose, including
commercially, **as long as you credit Kakuto with a link to [kakuto.io](https://kakuto.io)** wherever they are used,
e.g. on the page or screen that uses the model, or in your credits or about page:
> Fighter model by Kakuto (https://kakuto.io)
In a website or app, make "Kakuto" or the address a link to https://kakuto.io. If you changed the files, say so
("Based on the fighter model by Kakuto (https://kakuto.io), modified"). Full terms: [`LICENSE`](LICENSE).
To cite it:
```bibtex
@misc{kakuto_fighter_2026,
author = {{Kakuto Technology GK}},
title = {Kakuto fighter},
year = {2026},
howpublished = {\url{https://huggingface.co/kakuto-io/fighter}},
note = {https://kakuto.io}
}
```
MuJoCo's engine (`web/mujoco-3.16.0.wasm.gz`) is not covered: it is © DeepMind Technologies Limited, under the Apache
License 2.0 (`web/MUJOCO_LICENSE`).
|