Abstract
The design of complex self-organising systems producing life-like phenomena is a central goal of artificial life. Lenia, a family of continuous cellular automata (CA), generates a wide diversity of self-organizing patterns, yet finding stable "creatures" is difficult, and global rules prevent multi-species interaction. This paper introduces Flow-Lenia, a mass-conservative extension that solves these issues through mass conservation and parameter localization. We demonstrate its effectiveness in generating complex behaviors and multi-species simulations where locally coherent update rules define the properties of emerging creatures.
Introduction
The quest for open-ended evolution in artificial life often centers on cellular automata (CA). Lenia, a continuous generalization of Conway’s Game of Life, has shown promise in generating lifelike "creatures" or spatially localized patterns (SLPs). However, Lenia suffers from two primary limitations: the "needle in a haystack" problem, where viable parameters are extremely rare, and the "global physics constraint," which prevents different species from coexisting in the same simulation.
Methodology: Mass Conservation and Parameter Localization
Flow-Lenia addresses these challenges by transforming the standard CA update rule into a fluid-like flow system.
Mass Conservation via Reintegration Tracking
Unlike standard Lenia, where cell states can grow or decay arbitrarily, Flow-Lenia treats activations as physical matter. The system calculates a flow vector toward areas of high "affinity" (growth), using a method called Reintegration Tracking (RT). This ensures that total mass is conserved, preventing the "explosive" growth common in standard CA and making stable SLPs the default state.
Parameter Localization
To enable multi-species interactions, Flow-Lenia embeds update rule parameters—effectively a "genome"—directly into the CA state. As matter flows, its specific parameters move with it. When different parameter sets meet, they can be blended via weighted averaging or selected via softmax sampling.
Key Insights: Emergent Symbiosis and Metabolism
By localizing parameters, the researchers enabled the study of multi-species dynamics. In large-scale simulations, they observed emergent symbiosis, where distinct species (parameter sets) formed cooperative structures, such as one species acting as a protective membrane for another.
Furthermore, by introducing a "food" map and a constant decay rate, the system creates an intrinsic selective pressure. Creatures must actively navigate toward resources to maintain their mass, providing a natural basis for evolution without external fitness functions.
Results and Discussion
Flow-Lenia demonstrates significantly higher stability than its predecessors. Optimization for tasks like directed motion consistently succeeds in Flow-Lenia, whereas standard Lenia often results in unstable patterns.
The framework discovered a wide variety of complex forms, including "snakes," gyrating dots, and organoid-like structures. Evolved creatures also displayed robust navigation capabilities, such as climbing chemical gradients (chemotaxis) and avoiding obstacles.
While Flow-Lenia marks a significant step toward open-ended evolution, challenges remain in mathematically defining individual "agents" within these emergent systems and achieving unbounded increases in complexity. Nonetheless, the integration of mass conservation and localized genomes provides a powerful new toolkit for artificial life research.
Summary updated