Neural CA for Self-Assembly – Decentralized, Self-Repairing (0.003 MSE) (github.com)

Researchers released a compact Neural Cellular Automata (Neural CA) system that learns a single local neural rule to self-assemble and fully self-repair emergent structures. Each cell only observes a 3x3 neighborhood and there is no central controller; yet from a single seed the identical local policy grows target patterns, withstands extreme damage (the "Wolverine" training protocol) and achieves 100% recovery with a very low training loss (0.003 MSE). The trained model is tiny (~20 KB), trains in about 5 minutes on an RTX 3060, runs in pure NumPy (no GPU required for inference) and is released with demo and training scripts under GPLv3. Technically, the system uses a small CNN with residual connections, stochastic update scheduling, gradient clipping and value clamping for stability, and was trained in PyTorch (Conv2d + autograd). The simulator includes simple physics (gravity, collisions, magnetic linking) and an agent-based M-Block–style cube model to validate distributed self-organization. Key implications: practical, embeddable controllers for modular robotics and swarm systems, resilient programmable-matter behaviors on constrained hardware (even microcontrollers), and a concrete step toward fault-tolerant distributed systems where global intelligence emerges from tiny identical local networks. Code and demos are available in the neural-automata-lab repository.