Two hundred identical cubes guess together what shape they make

Sakana AI’s lab has 200 autonomous cubes work together to identify their shared shape—a major breakthrough for connected devices.

Sakana AI, along with the IT University of Copenhagen and Autodesk, has published a paper in Nature Communications on cubic blocks capable of collectively recognizing the shape they form, without any central coordinator.

The setup is spartan. Each block is a small cubic printed circuit board, equipped with connectors on its six faces, a microcontroller, and an LED that displays its current hypothesis. They all run the same neural network and communicate only with their immediate neighbors. None of them knows its own position, and none knows which object it belongs to. Yet, through purely local exchanges, the assembly converges on a common answer: airplane, guitar, boat, table.

The method relies on neural cellular automata—local update rules that are learned rather than hand-coded—extended here to the third dimension and embedded in hardware. Across nearly two hundred physical blocks, the collective reaches agreement in fewer than sixty cycles, or about three minutes. The researchers report that the system tolerates the failure of some of its modules, except when the shape features a bottleneck: in a guitar, a single failure on the neck cuts the object in two and causes the recognition to fail.

The most intriguing aspect lies in what emerges along the way. By inspecting the network's internal channels, the authors observed activation patterns reminiscent of morphogens, the diffusing molecules that inform embryo cells of their position. A left-right axis, a radial gradient, and then, to distinguish a chair from a table, an anteroposterior axis that propagates from the backrest and tips the collective toward the correct answer. Finally, the approach extends to self-repair: trained to detect a missing neighbor, the blocks can guide their own reconstruction, adding modules until no further damage is detected.