iiinova — where your digital self keeps evolving

Manuel Baltieri @manuelbaltieri.bsky.social · 19/09/2026

Conclusion: can PP, FEP and/or active inference be used to motivate biological naturalism? No more and no less than Kalman filters, Bayes theorem and/or PID controllers can. doi.org/10.1017/S014... Preprint: osf.io/preprints/ps... 12/12

doi.org

Steam-engine naturalism | Behavioral and Brain Sciences | Cambridge Core

Steam-engine naturalism - Volume 49

010

Manuel Baltieri @manuelbaltieri.bsky.social · 19/09/2026

We then propose three moves: - push for biological naturalism as is, with tools other than PP, FEP and/or active inference - drop biological naturalism - take parts of biological naturalism as a new claim, and characterise its universal (biological) properties. 11/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 19/09/2026

These can all be found with a quick search on google scholar and more will appear, I'm sure, after 2022. So, plenty of models of non-biological systems. Obligatory at this point: 10/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 19/09/2026

More evidence: from my last presentation in 2022 on our previous BBS work, doi.org/10.1017/S014.... "Markov blankets of life, mind, self, sex and gender, pain experience, religious practices, climate and ecosystems, social systems, cultures, cryptos, quantum systems, …" 9/

doi.org

The Emperor's New Markov Blankets | Behavioral and Brain Sciences | Cambridge Core

The Emperor's New Markov Blankets - Volume 45

100

Manuel Baltieri @manuelbaltieri.bsky.social · 19/09/2026

(Some) Evidence: - PID control as active inference mdpi.com/1099-4300/21... - Watt governor as active inference direct.mit.edu/isal/proceed... - active inference for robotics arxiv.org/abs/2112.01871 8/

mdpi.com

100

Manuel Baltieri @manuelbaltieri.bsky.social · 19/09/2026

Now to the issues: Can PP, FEP and active inference be used to model aspects of biological systems? Yes. Are they very successful? Not always, see for instance philosophymindscience.org/index.php/ph.... Do they model exclusively aspects of biological systems? No. See below. 7/

philosophymindscience.org

Laying down a forking path: Tensions between enaction and the free energy principle | Philosophy and the Mind Sciences

Philosophy and the Mind Sciences (PhiMiSci) focuses on the interface between philosophy of mind, psychology, and cognitive neuroscience. PhiMiSci is a peer-reviewed, not-for-profit open-access journal...

110

Manuel Baltieri @manuelbaltieri.bsky.social · 19/09/2026

3. Active inference was initially proposed as the action-oriented arm of the FEP, or as a complementary view of PP focusing on action and motor control with roots in cybernetics and control theory. See Anil's: open-mind.net/papers/the-c... 6/

open-mind.net

The Cybernetic Bayesian Brain — Open MINDOpenMIND

100

Manuel Baltieri @manuelbaltieri.bsky.social · 19/09/2026

2. The free energy principle was initially proposed a theory of brain function, later as theory of the origins of life, and then as an extension of studies of physical systems (biological or not) in the tradition of Jaynes. See next. 4/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 19/09/2026

1. Predictive processing has historically developed as a framework that adapts ideas from Kalman filtering to study the temporal and hierarchical structure of brain functions. See: sciencedirect.com/science/arti... 3/

sciencedirect.com

100

Manuel Baltieri @manuelbaltieri.bsky.social · 19/09/2026

In our commentary, w/ @kanair.bsky.social, we build our argument on a very brief historical overview of a body of work that already displays, IMO, how PP, FEP and active inference, taken individually or together, don't say much that is exclusive to biological systems. 2/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 19/09/2026

Happy to see this coming together. Great work by Anil, pushing his ideas out there to foster a much needed discussion. I remain unconvinced by any of Anil's arguments in favour of "biological naturalism", and yet see no good evidence at the moment for machine consciousness. 1/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

In other words: which differences (do not) make a difference for the agent. Paper: manuelbaltieri.com/assets/pdf/B... For some related ideas see also: Ay & Löhr (2015), "The Umwelt of an embodied agent - a measure-theoretic definition". 17/17

manuelbaltieri.com

000

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

Overall, we propose an operationalisation of "habitat" as a structure-preserving abstraction of an environment, relative to a class of embodied agents. Here, morphology, sensing, dynamics and actions determine which environmental differences remain behaviourally available. 16/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

These three are only candidate coarse-grainings. Turning them into quotient POMDPs would require full observation and transition maps, with each proposed partition preserved under every admissible action. 15/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

- a Braitenberg crab, embodying lateral motion because only a distinction on a source bein left/centre/right remains. 14/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

- a Braitenberg sea urchin, with only radial symmetry so that only distance to the source matters, and 13/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

We then sketch three candidate habitats in the same light field: - a Braitenberg sunflower, pinned in place, its bearing remains but distance to the source stops mattering, 12/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

This is because the orbit map, q₂b((x,y), θ) = (r, θ_rel), induces a bisimulation quotient. The habitat removes absolute position around the source and retains distance and signed relative heading: how far away the source is, and how much the vehicle must turn to face it. 11/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

Now, place the vehicle in two configurations related by rotation around an isotropic light source, with distance and relative heading fixed. The sensors give the same readings. Under any fixed wheel-speed action, the successors remain related by the same rotation. 10/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

To show this in action, we consider a class of Braitenberg vehicles. Vehicle 2b has two light sensors and two wheels, wired contralaterally. Light on the left accelerates the right wheel, turning the vehicle toward the source; equal readings drive it forward. 9/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

The quotient S/~ forms the habitat’s state space. Its observations and transitions are inherited from the environment. This construction is relative to a chosen agent-environment boundary. Different interfaces or bisimulation relations can produce different habitats. 8/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

If s~s′, then: (i) they induce the same observation law, and (ii) for every admissible action, they assign the same transition probability to every equivalence class. The quotient preserves observable dynamics. 7/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

This coarse-graining identifies environmental states that are indistinguishable from the perspective of an agent type. We define it using a standard bisimulation equivalence. See our previous bsky.app/profile/fros.... 6/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

We model the environment as a POMDP Q=(S,A,T,O,M). Its states describe the physical setup, actions and observations define the interface through which a controller can affect and sense it. The habitat is another POMDP, obtained by an appropriate coarse-graining of states S. 5/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

The habitat is shaped by morphology shared among members of a species and by their sensorimotor interface. The more familiar notion of umwelt is instead an individually enacted world, shaped by policy, learning and sensorimotor history. 4/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

To clarify this, we follow Baggs & Chemero (2021), "Radical embodiment in two directions", differentiating between: - a physical environment, and - a habitat, the “effective physics” of an agent type. Importantly, this work makes a distinction between habitat and umwelt. 3/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

Consider how humans and insects can share the same environment while registering different parts of it. Humans see the familiar visible spectrum, while some insects also sense ultraviolet light. So, which environmental distinctions survive at an agent’s interface? 2/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 13/07/2026

This time, celebrating 19 days of being “on hold” on arXiv! “A Braitenberg vehicle’s habitat”, w/ @frosas.bsky.social and Filippo Torresan. Accepted at ALIFE 2026. What part of the environment constitutes an agent’s effective world, its habitat? 1/

181

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

Full paper: www.sciencedirect.com/science/arti... Python code: github.com/FilConscious... The original draft has been greatly improved thanks to the helpful feedback of reviewers and colleagues. Thank you very much for your time and effort! 18/18

sciencedirect.com

000

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

TL;DR: Active inference agents can learn without knowing performed actions but they pay a computational price and require careful design to overcome some learning issues. Action information (efference copy) is not necessary, but it certainly provides a big advantage. 17/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

Future work could examine if action-unaware agents encounter similar problems when they need to learn the observation matrices instead. It would be also worth looking into a more principled way to make these agents work and that could scale to more complex scenarios. 16/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

This raises a fundamental question: which formulation is more biologically plausible? Action-unaware fits the proprioceptive prediction narrative better, but action-aware is computationally and algorithmically more tractable. 15/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

...both tweaks to the algorithm can be removed (this was done at episode 125 in the figure) and action-unaware agents can engage in retrospective policy inference. This is only problematic when transition matrices are being learned. 14/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

As a further result, we show in the appendix that, once an action-unaware agent has learned the environment dynamics, it can engage in policy inference at each step of an episode without dramatic performance drops... 13/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

Why does the fix work? By committing to one policy early (low entropy on policy distribution), action-unaware agents gain certainty about what action was likely taken at each step. They can then correctly attribute observed transitions to the right action matrix. 12/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

Our second finding is that with a couple of adjustments, (1) limiting policy inference to the first step of every episode and (2) a probability bonus mechanism for policy selection, action-unaware agents can learn and nearly match the performance of action-aware agents. 11/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

Our first key finding is that vanilla action-unaware agents fail to solve both navigation tasks. Without knowing which action they executed, they can't correctly attribute state transitions to actions so they are unable to learn the B matrices. 10/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

This makes learning the action-dependent transition matrices, or B matrices for short, much harder in action-unaware agents than action-aware ones because the former must infer what action caused each state change in an episode. 9/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

Concretely, action-aware agents evaluate policies based on future actions only whereas action-unaware consider the entire past, present, and future trajectory. That is, in the former case policy inference is only prospective whereas in the latter is also retrospective. 8/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

We built a Python implementation comparing action-aware vs action-unaware agents in simple navigation tasks (T-maze, 3×3 grid world). Both use the same active inference architecture for discrete state-spaces. The only difference is whether executed actions are known. 7/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

This choice reflects a long-standing debate in motor control. Some frameworks (e.g., equilibrium point hypothesis) claim humans don't have or need explicit access to motor commands. Others (e.g., optimal motor control) assume past action information is available for planning. 6/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

We compared two different implementations of active inference agents that differ in whether at the perceptual inference and learning stages executed actions are available (action-aware) or not (action-unaware). 5/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

Do active inference agents need to have access to past actions (cf. efference copy)? Or can they as well perform successfully in a task by inferring their motor trajectory from observations alone? 4/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

Active inference views agents as implementing approximate Bayesian inference via free energy minimisation: this is central for inferring the current state from observation, planning, selecting an action, and learning a model of the environment. 3/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

In active inference, the answer is not clear cut, as agents have been formally described with and without access to executed actions. The implications of this assumption have not been systematically examined before and we set out to do precisely so in our latest article. 2/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 10/07/2026

New work led by Filippo out on Neurocomputing! W/ @ksks.bsky.social , @kanair.bsky.social and me. Does an agent need to be "aware" of its executed actions to learn to achieve a certain goal? In reinforcement learning, the answer is generally "yes". 1/

151

Manuel Baltieri @manuelbaltieri.bsky.social · 02/07/2026

Coalgebraic determinisation is thus broader than NFA -> DFA. With different choices of parameters (functor type, monad, and algebraic structure), it gives powerset construction, totalisation and, as we show here, Bayesian filtering updates. Paper: arxiv.org/abs/2607.00034 18/18

arxiv.org

Bayesian updates from coalgebraic determinisation

The powerset construction is the classical determinisation procedure for nondeterministic finite automata. In the coalgebraic setting, this construction has been generalised to structured coalgebras, ...

000

Manuel Baltieri @manuelbaltieri.bsky.social · 02/07/2026

This gives a different final coalgebra semantics, i.e., for all n>0, causal (non-anticipatory) stochastic behaviours b_n : I^n -> D(O^n) compatible under prefixes. These give distributions of output *sequences*, with outputs depending only on inputs available up to time n. 17/

100

Manuel Baltieri @manuelbaltieri.bsky.social · 02/07/2026

Main result: For T = D and an algebra structure given in the paper, coalgebraic determinisation of supported Mealy machines is exactly unifilarisation. The determinised state space is D(S), and the transition map of the determinised coalgebra is the Bayesian filter update. 16/

100