Neural Particle Automata: teaching particles to self-organise
Researchers trained particles to form complex shapes without central control. Each particle runs the same neural network, learns local rules, and the swarm organises itself.
I spent twenty minutes watching Neural Particle Automata simulations this morning. Particles that start scattered randomly learn to arrange themselves into shapes like letters, spirals, or organisms. The idea is simple. Each particle runs the same small neural network. It sees the positions of nearby particles and outputs a force vector. No central controller. No blueprint. Just local interactions. The network trains by gradient descent. You define a target shape as a loss function, usually the distance between particle positions and the shape you want. The network learns rules that make particles move toward that configuration. Once trained, the particles maintain the shape even when you perturb them. What caught me is how cheap this is. The demo runs in a browser. Each particle’s network has maybe 32 to 128 parameters. You can train a swarm of 200 particles in a few minutes on a laptop. Compare that to diffusion models where inference alone costs dollars per image. The failure modes are interesting. Some target shapes never converge. Particles cluster in the wrong spots or oscillate forever. The paper does not explain why certain geometries are harder, just shows which ones worked. That is typical for self-organising systems: you get emergence but not guarantees. I see two directions. One is using this for procedural content in games. Train a particle system once, ship the tiny network, let it reconstruct the shape at runtime. The other is swarm robotics. If you can encode formation rules in a 5KB model, you can distribute it to cheap hardware. The code is open. The models are small enough to inspect by hand. I might fork it and see if particles can learn to avoid obstacles while forming shapes. That loss function would be messier but the compute cost is low enough to experiment.