Nouli Doom: ViZDoom corridor control

Comparison with Jev

Recorded comparison with Jev 1.14.0, dated 31 October 2026.

Collect the green armor and finish alive in Deadly Corridor. Nouli Doom and the recorded Jev setup start from the same game state with the same structured-state and goal interface.

In this recorded episode, Nouli collects the armor and survives; Jev's run ends before collecting it.

Playback follows simulator time; a panel holds its last available frame when its recording ends. This example was selected after scoring and illustrates this recorded setup, not a general ranking of the models.

Recorded evaluation · Setup and limitations.

Model at a glance

Property	Value
Parameters	23,435
Input	42-float structured state; 54 legal joint actions
Tested runtime	macOS and Linux ARM64 / CPU
License	Apache-2.0; upstream notices apply

Quickstart with Nouli CLI

Requires Python 3.12, the HF CLI, and Nouli CLI 0.4.2. Hugging Face hosts the files; inference runs on your machine. The Python server loads the weights, and Nouli runs the supplied decision workflow. No hosted API key is required.

In the first terminal:

Terminal 1 · prepare and serve

hf download nouli-ai/nouli-doom --revision v1.0.0 --local-dir ./nouli-doom
cd nouli-doom
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python prepare.py --input-json examples/input.json
python serve.py --device cpu

Once the server reports ready, open a second terminal in nouli-doom:

Terminal 2 · run the workflow

nouli --config nouli/nouli.toml --store ./.nouli-store \
  workflow run nouli/example-workflow.json --json

This example returns one action for the supplied state. Running an episode also requires the simulator loop to apply actions and provide each next state.

Base models and training data

Student: a PPO multilayer perceptron.
Teacher: fabiononato/doomfly-rl, checkpoint malecns49k_noconn.

Training data: Teacher trajectories, DAgger and simulator reinforcement learning.

Trained through imitation learning from the public Doomfly teacher and simulator reinforcement learning. Doomfly supplied labels; its weights are not the student's direct initialization. Source revisions and provenance.

Limitations

Validated on Deadly Corridor at skill 5 with the supplied state/action wrapper.
Inputs include structured state and goal assistance; these results do not establish pixel-only or general Doom performance.

License and citation

Apache-2.0, with upstream notices and dataset attribution retained.

Nouli AI. Nouli Doom. Cite the published model version and commit.