This document is a stable routing note for package users. It intentionally does
not maintain a full inventory of implemented modules, examples, or training
surfaces. Those details change quickly while state_collapser is still
pre-alpha, and duplicating them here would make this file drift from the repo.
- New to the package: usage overview
- Trying to understand the tower runtime: tower runtime mental model
- Trying to train with your own learner: training surface quickstart
- Working with freeze-and-lift training: fiber-conditioned training
- Using your own training loop: own training loop guide
- Adding a tensor/model boundary: tensorization boundary
- Integrating a Gymnasium environment: Gymnasium integration
- Looking at downstream applications: downstream applications
- Looking for exact provisional API surfaces: API notes
state_collapser is an installable Python package with real implemented runtime,
example, training, and benchmark surfaces. It is still pre-alpha, so users should
distinguish between:
- stable top-level package metadata
- documented provisional engineer-facing surfaces
- importable internal modules that may still change
The stable top-level import surface remains intentionally small:
import state_collapser
print(state_collapser.__version__)For public API expectations, read public API policy.
Install the RL-facing optional dependency group with:
pip install -e ".[rl]"Install the model-backend optional dependency group with:
pip install -e ".[ml]"The backend-independent tensorization layer lives in
state_collapser.training.linearization and is documented through the training
package. It does not import Torch at module import time. The ml extra is needed
for state_collapser.training.torch and its Torch batch-conversion helpers.
Install everything currently defined for local development with:
pip install -e ".[dev,rl,ml]"ROS 2 is intentionally not represented as a normal Python optional dependency.
It is a broader systems integration boundary whose concrete attachment points
should be designed explicitly rather than implied by a PyPI extra alone.
Use the README for the project overview, docs/usage for engineer-facing
workflow guidance, and docs/api_notes for exact provisional surfaces. Treat
this document as a map, not as the implementation inventory.
If you are maintaining a downstream package such as HGraphML, start with the
downstream-applications note and the public API policy before depending on new
submodules.