Python developers, researchers, and engineers implementing continuous-state Active Inference models.
A tested, documented continuous-state Active Inference library in Python.
Lead: Daniel Corva
cpomdp is an open-source Python library for continuous-state Active Inference. It provides factor-graph message passing, Kalman and reference filtering, state-dependent observation and process noise, certifiable tolerances, and matched classical control baselines.
cpomdp brings rigorous continuous-state Active Inference to Python. Agents maintain Gaussian beliefs over continuous latent states, update via message passing on Forney-style factor graphs, and select actions by minimizing expected free energy over declared action sets with certified warrant tolerances.
Contributors interested in numerical analysis, JAX/NumPy reference kernels, Kalman filtering, and continuous generative modeling are welcome to participate via Discord and GitHub.
The library is developed in the open: source on GitHub, documentation at cpomdp.inferogenesis.com, and releases published to PyPI as cpomdp. The first associated paper, "State-Dependent Observation Noise Reintroduces Epistemic Value in Linear-Gaussian Active Inference" (arXiv:2607.20306, 2026), sets out the result the library's state-dependent noise handling implements.
Related resources
Audience: Newcomer
Primary public community channel for conversation, coordination, and newcomer orientation.
Audience: Researcher
Public ecosystem shortlink for Institute context, projects, activities, and conceptual maps.
Audience: Developer
Source repository for the cpomdp continuous-state Active Inference library.
Audience: Developer
Documentation for the cpomdp library: API reference, tutorials, and worked examples.
Audience: Developer
Released cpomdp packages on the Python Package Index, installable with pip.
Audience: Researcher
Preprint showing that state-dependent observation noise reintroduces epistemic value in linear-Gaussian Active Inference; the result cpomdp implements.