Five brain circuits say embodied AI should build world models from action first, language second

Grounded world models in biological organisms and future embodied AI

Giovanni Pezzulo, Davide Nuzzi, Marco D'Alessandro, Riccardo Proietti, Roberto Bottini, Paul Cisek

q-bio.NC, cs.AI

2026-07-15

Five neural circuits anchor this argument: biology builds grounded world models from embodied action first, attaching language on top — the reverse of language-first embodied AI.

What problem this solves

Getting a robot to understand "pass me the cup, but it's fragile" and act on it takes more than language. Today's embodied AI, exemplified by visuo-language-action (VLA) models, first pretrains a large language model on massive text, then attaches visual, proprioceptive and action information onto that linguistic scaffold. The recipe is well-defined engineering, but it is passive training: data is selected and curated externally, the agent does little of the gathering, and language stays dominant throughout.

The authors' central question is whether this "language as scaffold" design has the order backwards. Neuroscience and developmental psychology paint the opposite picture: organisms first grow a "grounded" world model through sustained interaction with the environment, and language is attached later. The paper builds the case with five concrete neural circuits and asks whether these biological principles are essential for future embodied AI or merely incidental.

Method

This is a perspective, not a methods paper: no new model, no new experiment. The approach is to use five well-validated neural circuits as case studies and extract shared principles of grounded world modelling.

Spatial and conceptual navigation (hippocampal-entorhinal system). Place cells mark specific locations; grid cells build a toroidal-manifold coordinate system used for path integration. The system is intrinsically preconfigured: grid-cell firing and spontaneous hippocampal dynamics take shape before the animal has much navigational experience, then get aligned with the external world through action. The same circuit is reused for abstract spaces. When macaques choose in a value space defined by reward magnitude and probability, and when humans reason in a social space of competence and popularity, fMRI shows the six-fold symmetric, grid-like code. Conceptual thought reads as internalized navigation.
Affordances and object interaction (dorsal visual stream). Perception's main job is not to describe the world but to find out what can be done, the affordances. The dorsal stream lays out affordances; the ventral stream and frontal cortex select among them, and selection is a continuous competition. The circuit also predicts future affordances, so sequential decision-making becomes navigation across a landscape of present and future affordances. Motor-control circuits can be detached offline and reused to support imagination and reading others' intentions.
Active perception and exploratory learning (neuromodulatory control). Organisms gather experience through active interaction, not passive exposure. Systems such as the noradrenergic one maintain a "meta" world model that estimates expected information gain and toggles between exploration and exploitation accordingly. The key point is that the system keeps an explicit estimate of its own epistemic uncertainty and knowledge gaps.
Allostatic control, emotion and interoception. The brain regulates the body's internal milieu anticipatorily (allostasis), finding water before thirst hits. The system infers latent somatic states from interoceptive signals, and emotions such as joy or sadness are read as inferred interoceptive states. It also supplies intrinsic good/bad value signals that come from the organism's own dynamics rather than an externally imposed reward function. Needs and goals are present from the outset instead of added later.
Self-other distinction (corollary discharge / efference copy). When a motor command is issued, a copy of it (the efference copy) is sent to sensory systems to attenuate predictable self-generated input, highlighting what likely comes from outside. This is the basis for telling apart "I caused this" from "the world did", and for the sense of agency.

Collapsing the five circuits yields four recurring themes: intrinsic dynamics as the starting point for learning; action as the means of aligning those dynamics with the world; autonomous, open-ended experience preferred over passive assimilation of external data; and early predictive and control mechanisms scaffolding higher cognition such as reasoning, planning, imagination and language.

Results

A disclaimer up front: this paper runs no AI system and offers no method-versus-method baseline. It assembles established neuroscience to support a position, so the "results" below are facts about brains, not benchmark numbers.

Neural circuit	Representative finding	Lesson for embodied AI
Hippocampal-entorhinal	Grid cells form a toroidal topology; six-fold grid-like code in value and social spaces	World models can be intrinsically preconfigured, then aligned by action
Dorsal/ventral visual streams	Competition among affordances; motor circuits reused offline	Perception serves action; same circuits support imagination and theory of mind
Noradrenergic modulation	Explicit estimate of expected information gain; explore/exploit control	Agents should track their own uncertainty explicitly
Allostatic-interoceptive	Anticipatory regulation; intrinsic value from internal dynamics	Goals should be intrinsic to the system, not imposed later
Corollary discharge	Cancels self-generated sensory input; separates self from world	Distinguish "I caused it" from "the world did"

One often-cited minimal example: C. elegans has fewer than 400 neurons, yet its whole-brain dynamics spontaneously organize into a preconfigured locomotor manifold (seven fictive locomotor states) recoverable from spontaneous activity alone, with no real movement or environmental feedback. The authors treat it as a template for a minimal grounded model, where internal dynamics mostly encode the organism's own action repertoire rather than the fine structure of the external world.

Why it matters

For people building embodied AI, the value here is not a runnable recipe but a directional argument: make training active, intrinsically driven, continual and social, rather than piling on more language data. It connects to several live threads: the world-model line of LeCun, Hafner's Dreamer series, and Ha and Schmidhuber; Silver and Sutton's "era of experience"; and Friston's active inference.

The social-alignment angle deserves its own call-out. The paper argues that smooth, legible robot-human interaction needs more than fluent language: it needs action-mirroring circuits and non-verbal communication, which in the human brain are grounded in interoceptive and emotional systems. If future embodied agents are to genuinely respect human norms, these mechanisms may have to be written in as constraints or inductive biases from the start, not patched on late.

Honestly, this is a position paper, not a recipe. It reads as a research agenda: ammunition for those who believe scaled passive training will hit a wall, and an as-yet-unfalsified conjecture for those who do not.

Limitations

The authors are measured. They state plainly that none of these biological insights is currently central to generative or embodied AI, and that it remains hard to say whether they are essential or incidental; whether the passive paradigm will hit fundamental limits is unresolved; active and passive learning may converge to similar representations in domains such as space; and we still cannot tell which aspects of biological world-model acquisition are fundamental rather than contingent products of evolution. Engineering-wise, mainstays of current AI such as backpropagation, large-batch training and GPU parallelism may not transfer cleanly to a biological-style learning regime.

A few more doubts surface on close reading. First, the leap from "biology does this" to "AI should do this" is argued, not validated; nowhere does the paper show biologically inspired training actually beating scaled passive training. Second, the "language-as-scaffold is wrong" framing undersells what language pretraining buys in compositionality, world knowledge and transfer; the paper acknowledges language's power, but the headline stance is still a provocation. Third, the repeatedly cited findings such as the grid-cell torus and preplay come mostly from rodents and primates, and how to operationalize them for artificial agents is left open. Finally, a large share of the references comes from the authors' own group and the active-inference circle, a self-referential tendency that is natural for a perspective but worth bearing in mind.

Terms

World model: an internal model of the environment's latent states and action-dependent transitions, used to predict the consequences of one's own actions and of environmental dynamics.
Grounded: knowledge rooted in continuous, open-ended interaction with the environment rather than distilled from passive sensory or linguistic streams.
Affordance: an action possibility that environmental structure offers relative to an organism's bodily capabilities; it depends on both environment and body.
Grid cell / place cell: cells in the entorhinal cortex firing in a periodic hexagonal lattice, and cells in the hippocampus firing at specific locations, jointly forming a spatial "cognitive map".
Active inference: a framework unifying perception, action and learning as the minimization of prediction error.
Corollary discharge / efference copy: a copy of a motor command sent to sensory areas to cancel predictable self-generated input.

Source

What people are saying

Paper: Embodied AI Should Learn via Interaction, Not Passive Data Absorption AnnaCiaunica

Related papers

All paper explainers