Sakana: 197 electronic cubes identified their overall shape without a central computer

197 electronic cubes identified the shape of the whole assembly through messages from their neighbors

On July 13, a team from Sakana AI, IT University of Copenhagen, and Autodesk Research published a Nature Communications study of four assemblies containing 26–197 identical modules. In all three trials, the airplane, guitar, boat, and round table assemblies correctly identified their shapes within 23 cycles, or about 70 seconds.

Each module is a printed circuit board housed inside a cube. A microcontroller runs the same small neural network in every module, an LED displays the network's current answer, and connectors on all six faces transmit messages between neighboring cubes. The cube has neither coordinates nor a shared map of the assembly: it knows only its internal state and the messages received from cubes connected to it.

The authors trained a neural cellular automaton, a network in which every cube applies the same update rule. Each cube changes its internal state, receives new messages from its neighbors, and computes the next update. Signals propagate through the assembly until its different parts converge on the same label: airplane, guitar, boat, or table.

The network was first trained in simulation on seven object classes, after which the same rule was transferred to the physical modules. The authors found stable spatial gradients in its hidden channels: a cube's state varied with its location in the assembly. This signal gives each cube a relative sense of position, even though it receives no coordinates.

In simulation, the authors gave the network another task: identify the shape while also indicating one of six directions toward a missing cube. It identified the direction with a mean accuracy of 94.8% while maintaining 98.9% accuracy in shape recognition. The program then added a virtual cube in the indicated direction and repeated the calculation.

An assembly that needs to locate damage must first reach agreement about which object it forms. The team demonstrated this computational step using nearly two hundred cubes connected only to their nearest neighbors.