--- 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`).