Active inference is only as useful as the code that implements it. Across the ecosystem, a small number of toolchains carry most of the practical weight: Python packages for discrete-state simulation, a Julia framework for reactive message passing at scale, a decades-old MATLAB statistical platform that helped originate the theory, and symbolic-reasoning systems that pair free energy minimization with logical planning for robotics. This page surveys those toolchains, the Institute-affiliated development and learning activity around them, and the fragmentation that still separates them from a single coherent software stack.
Why the domain fits
Variational and expected free energy minimization are defined mathematically, but every application of active inference — from a robot arm to a psychiatric model to a policy simulation — ultimately runs on top of a generative-model implementation, an inference engine, and a numerical backend. Different toolchains have made different commitments here: discrete partially observable Markov decision processes versus continuous-state dynamics, batch computation versus reactive streaming, hand-specified generative models versus learned ones, and numerical versus symbolic representations of beliefs and goals. Because the Institute's other domain pages each describe active inference applied to a field, this page instead describes the shared computational layer underneath them — the toolchains a modeler actually picks up to go from theory to a working simulation.
Four toolchains, four paradigms
PyMDP is a Python package for simulating active inference agents in discrete-state Markov decision process environments, giving modelers a POMDP-style generative model, variational inference over hidden states, and expected-free-energy-based policy selection in a widely reused reference implementation. RxInfer.jl, developed at BIASlab in Eindhoven, takes a different route: it performs reactive message passing on Forney factor graphs rather than working directly with Bayes graphs or explicit POMDPs, so inference is event-driven rather than run on a fixed clock, with reported benefits for streaming data and for scaling to models with large numbers of parameters. SPM (Statistical Parametric Mapping), a MATLAB platform created by Karl Friston at the MRC Cyclotron Unit in the late 1980s for statistical analysis of fMRI, PET, and EEG data, predates active inference itself — its general linear model and Gaussian field theory machinery evolved alongside, and helped seed, the free energy formalism, and it remains in use today as both a historical cornerstone and a working analysis tool. Symbolic Cognitive Robotics, built around Jean-Francois Cloutier's karma_system project, takes a fourth path, combining symbolic reasoning with active inference so that generative models, beliefs, and goals are represented as symbolic structures rather than purely numerical ones, while still selecting actions by expected-free-energy minimization.
Institute-affiliated development and learning
The RxInfer.jl Learning Group is an active Institute project that collaborates directly with RxInfer.jl's developers on open-source development, including visualization techniques for the underlying factor graphs inside the code editor; the group has run regular synchronous sessions and drawn contributors with a range of Julia and Bayesian-modeling backgrounds working on applied projects. Symbolic Cognitive Robotics is the work of Jean-Francois Cloutier, a Research Fellow associated with the Institute, who has demonstrated the karma_system approach in robotic planning and symbolic problem-solving settings; robotics was also a dedicated focus of the Institute's second Applied Active Inference Symposium. PyMDP, while maintained outside the Institute, has been featured in the Institute's own ModelStream learning sessions, reflecting its role as a shared reference implementation across the wider community rather than a single-lab tool.
Open problems
The toolchains above do not share a common generative-model format, inference algorithm, or numerical backend, so a model built in one is not portable to another without substantial rewriting, and no standardized benchmark suite yet exists to compare them head-to-head on accuracy, speed, or scaling behavior. The discrete/continuous/symbolic split also means practitioners must commit early to a paradigm — POMDP-style discrete states, continuous message passing, or symbolic representations — with limited tooling for combining them within one model. Documentation and packaging maturity vary considerably: some tools have tutorials and structured learning groups, while others remain closer to research code than distributed software. As with active inference's other application domains, claims of adoption and performance advantage rest more on demonstrated proofs of concept and active community development than on systematic, independent comparison against established alternatives.
Reference Backbone
Karl J. Friston (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience. DOI: 10.1038/nrn2787. Christopher L. Buckley, Chang Sub Kim, Simon McGregor, Anil K. Seth (2017). The free energy principle for action and perception: A mathematical review. Journal of Mathematical Psychology. DOI: 10.1016/j.jmp.2017.09.004. Thomas Parr, Giovanni Pezzulo, Karl J. Friston (2022). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. MIT Press. Lancelot Da Costa, Thomas Parr, Noor Sajid, Sebastijan Veselic, Victorita Neacsu, Karl J. Friston (2020). Active inference on discrete state-spaces: A synthesis. Journal of Mathematical Psychology. DOI: 10.1016/j.jmp.2020.102447.