Bert de Vries

The result: a principled framework for embodied, agentic AI — robots, drones, and autonomous systems that perceive, learn, and act on-the-fly in real time.

Finally: Active Inference (AIF). AIF extends BML to embodied agents with a full commitment to variational inference for state estimation, learning, planning, and control. See t.ly/zcSdH

I recently had the pleasure to lecture at the Machine Learning Summer School in Melbourne t.ly/hX0Uo on Bayesian Machine Learning → Active Inference. All materials (slides + notebooks) available at github.com/bertdv/mlss-20… . Thread below 👇

This is what we ask ourselves in our latest paper (arxiv.org/abs/2508.02197), where we solve Minigrid environments with Active Inference. (Spoiler, it works!) Code available at github.com/biaslab/EFEasV… Done together with @mlmykola , @ThijsvdLaar and @bertdv0

What's even nicer: because our method injects priors locally, everything still works within @ReactiveBayes ' RxInfer.jl using message passing. Special thanks to my colleagues at @LazyDynamics for making this possible!

In Active Inference, a lot of time is spent on computing Expected Free Energy. What if we could tweak the generative model such that EFE can be minimised with traditional variational inference methods?

Backed Trojan Robotics (Team 24090) at the FIRST® Tech Challenge European Premier Event in Eindhoven (July 1–5)  . They hustled—coding, building, troubleshooting—and came away with 3rd in the Think Award. Proud to support their next steps. 🚀 #FTC #Robotics #STEM

@MovingFramesP @fchollet Technically, FEP leads to simultaneously active inference (updating states), active learning and active model selection. The term “active inference” is usually interpreted as an umbrella over all active free energy minimizing processes.

@fchollet Is active inference just inference or does it include training at the same time?

The Free Energy Principle is a pretty good idea, but its core value isn't to serve as a grand unifying "theory of everything" for cognition. Rather, its core insight is the rigorous emphasis on active inference, which has been badly missing from the deep learning era.

@fchollet x.com/bertdv0/status…

Some more refs: for implementation, rxinfer.com (toolbox), and arxiv.org/abs/2504.14898 (AIF planning as inference), and lazydynamics.com (company). Also check out noumenal.ai , stanhopeai.com, and verses.ai .

(6) Finally, FEP is more than a pretty good idea as it can be derived from first principles by information theory, see e.g., blog at t.ly/Bl2DO plus refs. An AIF process avoids ad hoc design choices often found in man-made AI algorithms.

Agreed with @fchollet on FEP (t.ly/s4gMV), but FEP is more than a pretty good idea, and there are more benefits to realizing an agent as an active inference (AIF) process beyond active data selection. I will mention a few below: