[Paper] Agnosiophobia in a virtual agent: behavioral and dynamical architecture in Lenia

Source: arXiv - 2605.30708v1

Overview

The paper explores how virtual agents—self‑organized patterns in the cellular‑automaton platform Lenia—behave when parts of their environment are rendered “invisible.” By deliberately occluding sensory information, the authors uncover a novel avoidance tendency they call agnosiophobia (fear of the unknown). This work bridges concepts from dynamical systems, artificial life, and robotics, offering fresh insight into how emergent agents might navigate informational gaps in real‑world settings.

Key Contributions

• Introduces agnosiophobia: the systematic avoidance of regions that provide no sensory data, extending the classic infotaxis framework.
• Sensory‑occlusion mapping: a quantitative method to probe each Lenia creature’s sensitivity to targeted blind spots.
• Dynamical‑systems interpretation: links avoidance behavior to the preservation of an agent’s morphological stability (its “shape attractor”).
• Demonstrates heading flexibility: shows that Lenia agents can reorient to maintain morphology, suggesting a higher‑level goal beyond simple foraging.
• Provides a reproducible experimental pipeline for testing information‑topography interactions in other emergent systems.

Methodology

1. Agent selection – A diverse set of pre‑evolved Lenia patterns (e.g., “Gliders,” “Spirals”) were used as test agents.
2. Environment design – The 2‑D Lenia world was divided into “visible” and “occluded” zones. Occlusion was implemented by zero‑ing the sensory field (the convolution kernel) over selected patches, so agents received no feedback from those areas.
3. Behavioral trials – Each agent was placed near an occluded region and allowed to run for a fixed number of update steps. Trajectories, heading changes, and morphological metrics (e.g., shape entropy) were recorded.
4. Sensitivity mapping – Systematically moved the occlusion patch across the environment and measured the resulting deviation in the agent’s path, producing a heat‑map of “fear zones.”
5. Dynamical analysis – The authors treated the agent’s state (position, heading, morphology) as a point in phase space and examined how occlusions perturbed its trajectory relative to its underlying attractor.

All steps were implemented in open‑source Lenia code, making the experiments fully reproducible.

Results & Findings

• Consistent avoidance – Across all tested patterns, agents steered clear of occluded patches, even when those patches contained no explicit hazards.
• Morphology preservation – Agents that entered occluded zones showed rapid deformation of their shape; those that avoided the zones maintained a stable morphology, suggesting a hidden objective of “shape integrity.”
• Heading flexibility as a strategy – Many agents performed smooth heading adjustments to circumnavigate blind spots, indicating an emergent planning capability.
• Sensitivity heat‑maps revealed distinct “fear profiles” for different pattern families, analogous to species‑specific risk maps in biological organisms.

In dynamical‑systems terms, occlusions acted as perturbations that pushed agents away from their morphological attractor; avoidance behavior can be seen as a self‑stabilizing feedback loop.

Practical Implications

• Robotics & autonomous navigation – Real‑world robots often encounter sensor drop‑outs (e.g., GPS loss, occluded LiDAR). Understanding agnosiophobic tendencies could inspire control policies that proactively avoid blind zones, improving safety.
• Design of adaptive AI agents – Embedding a “morphology‑preservation” objective may help AI systems maintain internal consistency when faced with incomplete data, useful for continual‑learning or self‑repairing software.
• Virtual world testing – Game developers can use Lenian‑style occlusion tests to evaluate NPC behavior under limited visibility, leading to more believable AI opponents.
• Explainable AI – Mapping sensitivity to information gaps offers a transparent way to diagnose why an agent makes certain decisions, aiding debugging and compliance.

Limitations & Future Work

• Domain specificity – Findings are based on a 2‑D cellular‑automaton; transfer to high‑dimensional physical robots remains an open question.
• Simplified sensory model – Occlusion was binary (on/off). Real sensors degrade gradually, so future work should explore graded information loss.
• Goal ambiguity – While morphology preservation appears to drive avoidance, the exact utility function is not formally defined.
• Scalability – Extending the sensitivity mapping to large‑scale, multi‑agent environments will require more efficient algorithms.

The authors suggest exploring richer sensory modalities, integrating learning mechanisms, and testing the concepts on embodied hardware as next steps.

Authors

• Jesse Cool
• Benedikt Hartl
• Michael Levin
• Samantha Petti

Paper Information

• arXiv ID: 2605.30708v1
• Categories: nlin.CG, cs.NE, nlin.AO
• Published: May 29, 2026
• PDF: Download PDF