Neural Particle Automata represent a shift from grid-based neural cellular automata, where individual particles move freely through space while following shared update rules. Each particle maintains continuous position and internal state, enabling complex morphological growth and pattern formation that traditional fixed-lattice approaches cannot achieve. This architecture allows emergent behaviors like self-healing, demonstrated dramatically when researchers cut particles from a simulated lizard that regenerated.

The system uses Smoothed Particle Hydrodynamics perception as its particle-based counterpart to convolutional perception. Rather than reading from fixed neighbors, particles aggregate nearby agents within a support radius using smooth kernels. These local measurements estimate density, smoothed state, density gradients, and moment matrices that capture neighborhood geometry. The approach preserves NCA locality while supporting irregular and dynamic particle configurations.

A demonstration visualizes these operators in 2D space, with particles represented as RGB-colored agents and kernel functions shown as distance-based plots. The moment matrix correction normalizes geometric distortion, making gradient estimates accurate for locally linear fields. Researchers highlight how the 0th-order and 1st-order state gradient estimators handle uneven sampling conditions.

This work opens possibilities for modeling biological systems, materials science, and distributed robotics where agents operate in continuous space rather than discrete grids. The self-healing capability suggests applications in resilient computing architectures and adaptive physical systems.