@sabaki-js/doom

Doom decisions (navigation + combat arbitration).

Questions: command, target, weapon. Work in progress: nav MiniLM (MAP01) + vision policy (arena).

Demo

npm run demo — the plugin decides on 3 canned MAP01 states (nano, no deps).

The mini model (MiniLM + exit compass) finishes MAP01 — 4/6 EXIT:

| mode | weights | closed loop | |---|---|---| | nano (TF-IDF, 112 terms) | src/weights.json | 0/6 | | mini (MiniLM ONNX, CPU) | sabaki-models | 2–4/6 |

Run your own demo

The classifier runs in Node, the game in Python (VizDoom). The JS server decides, the Python client plays — one HTTP call per game tick.

# 1. game + model stack (Python)
pip install vizdoom gymnasium pygame onnxruntime
git clone https://github.com/sabaki-js/doomLaya /tmp/doomlaya  # fork Apache-2.0 d'azalio/doomLaya, voir NOTICE

# 2. mini weights (~90 MB) + tokenizer
# from https://github.com/sabaki-js/sabaki-models/releases (doom_nav.onnx)
export ONNX=/path/to/doom_nav.onnx VOCAB=/path/to/tokenizer.json

# 3. JS decision server (needs sabaki@>=1.0.1 + onnxruntime-node)
npm i [email protected] onnxruntime-node
node serve_mini.mjs 8472

# 4. play one episode (second terminal)
DOOM_SERVE=http://localhost:8472 python examples/drive_serve.py 1

serve_doom.mjs serves the nano classifier instead (faster, weaker: 0/6 — it decides well on single states but cannot navigate a full level).

Usage

import { Sabaki } from "sabaki";
import plugin from "@sabaki-js/doom";

const sabaki = new Sabaki().register(plugin.default ?? plugin);
const r = sabaki.decide({ state: "..." });
// { domain: "doom", decision: ..., confidence: ..., plugin: "@sabaki-js/doom" }

Ask another question with input.question.

Part of sabaki — see the GUIDE.

License

MIT