NOULIComparison with Jev
Recorded comparison with Jev 1.14.0, dated 31 October 2026.
Open the side-by-side recording.
Move the T-shaped block into its target region. Nouli Push and the recorded Jev setup start from the same placement and receive structured keypoints and causal history.
In this recorded episode, Nouli completes the placement; Jev's run reaches the time limit. The unchanged DOT starter performed similarly under this controller, so the additional fine-tuning benefit remains uncertain.
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.
Model at a glance
| Property | Value |
|---|---|
| Parameters | 2.12 million |
| Input | Structured keypoints and causal state history |
| 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:
hf download nouli-ai/nouli-push --revision v1.0.0 --local-dir ./nouli-push
cd nouli-push
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-push:
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
Training data: lerobot/pusht_keypoints; 206 demonstrations / 25,650 frames.
Adapted from DOT using demonstration actions and an original-policy anchor. The starter already used these demonstrations; evaluation uses fresh simulator initializations. Source revisions and provenance.
Limitations
- Validated for numeric-keypoint control in PushT. Camera-input and real-world transfer are untested.
- Use the supplied state normalization, history and action-grid interface.
License and citation
Apache-2.0, with upstream notices and dataset attribution retained.
Nouli AI. Nouli Push. Cite the published model version and commit.