decart-robotics

The Python SDK for Decart's Oasis 3 Preview real-time world model, published to PyPI as decart-oasis.

pip install decart-oasis

from decart_oasis import A2VClient

with A2VClient() as client:          # set DECART_API_KEY in the environment
    client.prompt("driving in an urban area")
    result = client.infer([[0.2, 0.0], [0.2, 0.0], [0.2, 0.1], [0.2, 0.1]])

front = result.frames["front"]       # list of 4 RGB frames (H×W×3 uint8)

The SDK lives in sdk/ — see sdk/README.md for full usage and sdk/docs/python-sdk.mdx for the reference.

RL demo (oasis-demo)

extras/ holds oasis-demo: a depth-collision reward, a Gymnasium env wrapping the SDK, a Stable-Baselines3 PPO policy with behavior cloning, and a live notebook preview — everything on top of the SDK for training a driving agent in Oasis. It pulls the full ML stack (torch, stable-baselines3, transformers), so it's a separate package. See extras/README.md.

git clone https://github.com/DecartAI/decart-robotics.git
pip install -e decart-robotics/extras    # also pulls decart-oasis from PyPI

The end-to-end Colab training notebook is added in a later phase.

Development

uv sync                  # installs the SDK + oasis-demo (editable) + dev tools
uv run pytest            # sdk/tests + extras/tests
uv run ruff check .

Releasing

.github/workflows/publish.yml builds and publishes the SDK to PyPI via Trusted Publishing (OIDC) — triggered by a GitHub Release, or manually via Actions → Publish to PyPI → Run workflow (choose testpypi or pypi).