Sakana AI's cellular bricks self-classify their shape with no central controller

Sakana AI, with IT University of Copenhagen and Autodesk, ran nearly 200 physical bricks — each running an identical local neural cellular automaton, none knowing its own position — that reached 100% consensus on all four test shapes in under three minutes, tolerating 5% module dropout with little accuracy loss. The Nature Communications result moves collective-intelligence research out of simulation and onto hardware, showing decentralized self-classification survives noise and failure — a foundation for smart materials and reconfigurable robotics where global structure must emerge from purely local sensing.

Source: sakana.ai ↗