RL examples and notebook visualization for the Decart Oasis SDK — everything that is not the
core connection. It builds on the decart-oasis SDK (A2VClient) and adds:
oasis_demo.rl.reward— a depth-based collision-avoidance reward (DepthCollisionReward).oasis_demo.rl.env— a Gymnasium environment (OasisA2VEnv) wrapping the SDK client.oasis_demo.rl.sb3/rl.sft/rl.pretrained— a Stable-Baselines3 PPO policy, behavior cloning from recorded episodes, and a pretrained-policy loader.oasis_demo.live_preview/oasis_demo.visualization— a live notebook camera preview and GIF/H.264 encoders.
Because it pulls the full ML stack (torch, stable-baselines3, transformers, …), it is a separate package from the lightweight SDK.
git clone https://github.com/DecartAI/decart-robotics.git
pip install -e decart-robotics/extras # editable, so you can tweak the reward etc.oasis-demo declares decart-oasis (the SDK) as a dependency, so that single install also pulls the
published decart-oasis from PyPI — no separate step. Installing editable keeps the source under
extras/oasis_demo/ visible and modifiable — edit, for example, extras/oasis_demo/rl/reward.py,
restart the kernel, and re-run.
This repo is a uv workspace; from the repo root:
uv sync # installs decart-oasis + oasis-demo editable
uv run pytest extras/tests
uv run ruff check extras