🌀 nonlinear sciences

Agnosiophobia in a virtual agent: behavioral and dynamical architecture in Lenia

This paper introduces the concept of "agnosiophobia" in Lenia virtual agents, demonstrating that emergent patterns actively avoid sensory-deprived regions to preserve their morphology, thereby revealing how behavioral propensities interact with the informational topography of their environment.

Original authors: Jesse Cool, Benedikt Hartl, Michael Levin, Samantha Petti

Published 2026-06-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jesse Cool, Benedikt Hartl, Michael Levin, Samantha Petti

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world made entirely of digital pixels, where "creatures" aren't made of flesh and blood, but are self-sustaining patterns of light and energy. These creatures, called Lenia, live in a computer simulation. They don't have brains, eyes, or nervous systems. They are simply shapes that move, grow, and maintain their form based on a set of mathematical rules, much like a complex dance routine that repeats itself.

The researchers in this paper wanted to see if these digital shapes could "think" or react to their environment in a smart way, even without a brain.

The Experiment: Blinding the Creatures

To test this, the scientists created a special environment. Imagine walking through a foggy room where some parts of the floor are covered in thick, invisible tar. You can't see the tar, and if you step on it, you might trip or get stuck.

In the simulation, the scientists placed "blind spots" (regions where the creature cannot sense anything) in the creatures' path. These blind spots are like holes in the creature's vision. If a creature's "eye" (its sensory field) looks at a blind spot, it gets confused because it expects to see data but sees nothing.

The Discovery: "Fear of the Unknown"

The researchers expected these simple patterns to just crash into the blind spots or wander aimlessly. Instead, they found something surprising. Three out of the four creatures they tested developed a behavior the authors call agnosiophobia (literally, "fear of the unknown").

Here is how it worked:

The Reaction: As soon as a creature sensed that part of its view was blocked (the blind spot), it didn't just keep going. It turned away.
The Analogy: Imagine you are walking down a hallway. Suddenly, you feel a strange "static" or a lack of information coming from your left side. Even though you can't see a wall there, your body instinctively leans to the right to avoid whatever might be causing that weird feeling. These digital creatures did exactly that. They steered away from the "unknown" areas to protect their shape.

Why Did They Do It? (The Secret Sauce)

The creatures weren't programmed to avoid blind spots. There was no "if-then" rule telling them to turn. So, how did they do it?

The paper explains this using the concept of a Dynamical System. Think of a Lenia creature not as a solid object, but as a ball rolling inside a bowl.

The Bowl (The Attractor): The shape of the bowl represents the creature's "ideal" form. As long as the ball stays in the bowl, the creature is healthy and looks like itself.
The Walls (The Danger): If the ball rolls too close to the edge, it might fall out (the creature dies or explodes).
The Slope (The Turn): The sides of the bowl are slightly curved. When the creature hits a blind spot, it gets "pushed" toward the edge of the bowl.

Here is the clever part: The creatures that survived had a specific shape to their "bowl." The areas that would cause them to die were right next to areas that would just make them turn.

When the creature felt the "unknown" (the blind spot), the mathematical rules pushed it toward the edge of its safety zone.
Because of the shape of this zone, the only way to stay alive was to slide along the edge, which naturally caused the creature to turn.

It's like a skier who is about to fall off a cliff. The only way to stay on the mountain is to turn sharply. The skier doesn't "decide" to turn; the physics of the slope forces the turn to save them. These creatures turned to save their shape, and in doing so, they avoided the danger.

The Different Personalities

The researchers tested four different types of creatures, and they reacted differently based on their "body shapes":

The Orbium (O2u): The most skilled. It sensed the danger early and smoothly turned away, like a careful driver avoiding a pothole.
The K4s: This one was a bit clumsy. It would sometimes get so close to the danger that it would collapse into a different shape, spin around, and shoot off in the opposite direction.
The K6s: This one was a "skirt-walker." It would hug the edge of the blind spot, sliding along the boundary without turning away immediately.
The S1s: The unlucky one. Its "bowl" was shaped so that the danger was right at the front. It had no room to turn, so it just crashed and died.

The Big Picture

The main takeaway is that intelligence doesn't always require a brain.

These creatures didn't have a plan. They didn't know what a "blind spot" was. They simply had a set of rules that kept them alive. Because of the way those rules were built, avoiding the "unknown" was the only way to keep their shape intact.

The paper suggests that this is a fundamental way to understand how life works. Even simple systems, from single cells to complex robots, might have a "fear of the unknown" built into their very physics. They don't need to be told to avoid danger; if their structure is right, avoiding danger is just a natural side effect of trying to stay the same.

In short: To stay the same, you must sometimes move differently. These digital creatures moved away from the unknown not because they were scared, but because staying still would have meant falling apart.

Technical Summary: Agnosiophobia in a Virtual Agent

Problem Statement
The paper addresses the challenge of understanding behavioral propensities in novel, emergent agents that lack explicit cognitive architectures or evolved survival mechanisms. Specifically, it investigates how self-maintaining motile patterns in Lenia (a continuous cellular automaton) interact with the informational topography of their environment. The authors probe for "protocognition" by subjecting these agents to informational occlusions—regions of the grid from which no sensory data is available. The core question is whether these dynamic patterns exhibit adaptive behaviors, such as avoidance, when navigating environments where their "sensory field" (the convolution kernel) is partially blocked, despite having no explicit mechanism designed for such tasks.

Methodology
The study utilizes Lenia, a continuous cellular automaton where cells on a toroidal grid update their states based on a convolution of their neighborhood with a radially symmetric kernel ( $K$ ) and a growth function ( $G$ ).

Experimental Setup:
- Four stable, non-oscillating, motile creatures (O2u, K4s, K6s, and S1s) were selected from Chan's catalog.
- Occlusion Mechanism: The authors introduced a binary mask ( $B$ ) to the environment. In the update rule, occluded regions are excluded from the convolution operation. The potential $U_t$ is calculated by convolving the visible state ( $A_t \cdot (1-B)$ ) and renormalizing by the fraction of the kernel mass landing on visible territory. This simulates "blindness" to specific parts of the surroundings without altering the creature's internal rules.
- Testing Environments: Creatures were placed in 10 distinct environments with varying occlusion patterns (size, sparsity, convexity). Each creature started at the center facing 360 different orientations and ran for 2,000 steps.
Metrics and Analysis:
- Morphological Recovery: A Wasserstein-1 distance metric was constructed to compare grid states based on the distribution of active pixels (profiles). A creature is considered "recovered" if its state returns to a defined neighborhood (within a maximum distance $d_{max}$ ) of its canonical form.
- Outcomes: Trajectories were classified as recovery (return to neighborhood), death (mass drop < 0.01), metamorphosis, or explosion (unbounded growth).
- Sensitivity Mapping: To understand the mechanism, the authors performed targeted interventions, placing a persistent $3 \times 3$ occlusion at every non-zero pixel of a creature's body to map the resulting dynamics (frames to recovery, max distortion, and heading change).

Key Results

Emergence of Agnosiophobia: Three of the four creatures (O2u, K4s, K6s) exhibited a behavior termed agnosiophobia (fear of the unknown). They tended to avoid or reorient away from occluded regions, despite having no explicit "danger" representation.
- O2u (Orbium): Demonstrated the most robust avoidance, rotating away from occlusions detected on its flank.
- K4s: Often skirted occlusion edges or, upon frontal occlusion, transformed into an oscillating structure and re-emitted in the opposite direction.
- K6s: Exhibited "skirting" behavior, traveling along occlusion edges rather than turning immediately.
- S1s: Failed to exhibit avoidance; its lethal zones were located at its leading edge, causing immediate death upon encountering occlusions.
Spatial Structure of Sensitivity:
- Sensitivity maps revealed a gradient: regions sensitive to reorientation (causing heading changes) are spatially adjacent to lethal regions.
- Coupling of Fragility and Navigation: Large heading changes were only observed when the creature underwent long, morphologically distorted recovery paths. These paths correspond to trajectories that approach the basin boundary of the attractor. Fast, undistorted recoveries did not result in significant heading changes.
- S1s lacked a "buffer" zone; its lethal region preceded any reorientation zone, leading to immediate failure.
Dynamical Systems Interpretation:
- The authors interpret Lenia creatures as attractors within a high-dimensional state space.
- Free Variables: Due to the translational and rotational symmetry of the underlying rules, the creatures possess "free variables" (position and heading) that can change without altering the creature's morphological identity.
- Basin Geometry: Navigational competence emerges from the geometry of the basin of attraction. Perturbations that push the system near the basin boundary (but not outside it) allow the system to "offload" stress by shifting along the free variable (heading). This shift constitutes a reorientation.
- Partial Equifinality: The creatures exhibit equifinality regarding morphology (they return to their shape) but non-equifinality regarding heading (the final heading depends on the perturbation). This selective equifinality enables navigation.

Significance and Claims
The paper claims to illustrate the beginning of a roadmap for understanding how emergent agents' behavioral propensities interact with the informational topography of their world, not just the tangible one.

Protocognition: The observed agnosiophobia is not designed or evolved but emerges from the geometry of the system's attractor basin. It suggests that goal-directed behavior (persistence of form) can naturally give rise to adaptive navigation (avoiding information voids) through the coupling of perturbations to free variables.
Redefining Goal-Directedness: The authors challenge the restriction of goal-directedness to far-from-equilibrium systems with explicit energy variables. They propose that dynamical properties—such as attractor dimensionality, coupling to free variables, and basin boundary geometry—may be sufficient to derive a measure of goal-directedness. This allows for the characterization of unconventional intelligences (like Lenia creatures) on a continuum with biological agents.
Mechanism of Competence: The study demonstrates that navigational capacity is intrinsically linked to morphological fragility. The ability to change heading (navigate) requires the system to traverse states of high distortion near the boundary of survival.

Conclusion
The work establishes that simple dynamical systems can exhibit complex, adaptive behaviors like avoidance of unknown information through the interplay of symmetry, attractor geometry, and perturbation. The "fear of the unknown" is a consequence of the system's attempt to preserve its morphological attractor in the face of informational deprivation, utilizing its freedom to change heading as a mechanism for survival.