I'm 15 and built a Neural Cellular Automata system that self-assembles from a single seed and self-repairs from damage.
Each cell only sees its 8 neighbors and runs the same small neural network (20KB). No central controller. They coordinate into target shapes through local rules alone.
Key results:
• 0.003 MSE loss (near-perfect assembly)
• 100% self-repair from arbitrary damage
• Fully decentralized (no global communication)
• Trained in 5 minutes on consumer GPU
The repo includes training code, demos, and pretrained weights. Everything runs in pure NumPy.
This extends Google's 2020 Growing Neural CA work. I think it applies to modular robotics (MIT M-Blocks), programmable matter, and swarm systems.
Would love feedback on the approach or ideas for applications.
Each cell only sees its 8 neighbors and runs the same small neural network (20KB). No central controller. They coordinate into target shapes through local rules alone.
Key results: • 0.003 MSE loss (near-perfect assembly) • 100% self-repair from arbitrary damage • Fully decentralized (no global communication) • Trained in 5 minutes on consumer GPU
The repo includes training code, demos, and pretrained weights. Everything runs in pure NumPy.
This extends Google's 2020 Growing Neural CA work. I think it applies to modular robotics (MIT M-Blocks), programmable matter, and swarm systems.
Would love feedback on the approach or ideas for applications.