These identical bricks taught themselves to recognize their own shape. No central brain needed.

A recurring theme in Sakana AI's research is collective intelligence: sophisticated, robust behavior that emerges from many simple parts following local rules with no central controller. Until now, that line of work stayed in software. A paper published in Nature Communications marks the first time the same principles have been demonstrated in physical hardware, in a collaboration between researchers at IT University of Copenhagen, Sakana AI and Autodesk. The team is pushing the same idea that makes systems like robot swarms tick: local rules can produce global intelligence, see Xiaomi's own distributed factory robot.

The system is a collection of cellular bricks, each a small printed-circuit-board cube with electrical connectors on all six faces, a microcontroller, an LED to display its current class guess, and the electronics to power itself. Bricks stack together into arbitrary objects and communicate purely locally over a custom digital serial protocol. No brick is told its position or which shape it is part of. Each brick runs an identical neural cellular automaton (NCA), a deep neural network that learns the local update rules rather than having them hand-crafted, and iterates until the collective converges on a single shape label.

What the bricks do

The collective is tasked with distinguishing between objects that resemble planes, chairs, cars, tables, houses, guitars, and boats. Rather than matching a single predefined configuration, the system generalizes across entire classes of shapes. A five-legged table with legs at random positions was still classified as a table. A boat with its central bridge shifted off-center was still recognized. The network picks up abstract structural features, not memorized instances.

In simulation, the approach reached 98.97 percent accuracy. In hardware, the team built four distinct shapes ranging from 26 bricks (a guitar) to 197 (a round table). The bricks reached correct consensus on a plane, a guitar, a boat, and a table with a 100 percent success rate across all four, converging in fewer than 60 update cycles, roughly three minutes of real time. This kind of decentralized adaptation mirrors what researchers are now building into general-purpose robot controllers, see a new framework that lets robots adapt without retraining.

Damage detection and self-repair

Biological systems like planarians and axolotls regrow complex body parts through distributed cycles of sensing and regeneration with no central controller. The bricks were trained to detect whether a neighboring cell has been damaged, predicting either no damage or damage in one of six spatial directions. The system retained 98.9 percent shape classification accuracy while detecting damage with an average accuracy of 94.8 percent.

To test recovery, the team started from a small cluster of cells and repeatedly added new cells in the direction indicated by the existing ones until no more damage was detected. Without ever being trained to grow from only a few cells, the model recovered almost all shapes across all object classes with high accuracy. On larger simulated grids up to 64x64x64, the approach worked for shapes as complex as a fish, a Sakana AI logo, and a heart built from more than 18,000 cubes. The same principle of building robustness through local rules is gaining traction elsewhere, as seen in Mistral's acquisition of a physics AI firm to build digital twins.

Emerged communication strategies

The paper also examines how the collective actually solves the task. Early in the process, the hidden channels of the NCA establish left-right and radial activation patterns that resemble the morphogen gradients forming developmental axes in embryos. The same pattern explains how the system tells tables from chairs: for chairs, an anterior-posterior signal propagates from the backrest outward, gradually reclassifying cells that were initially read as "table" into a coherent "chair" consensus. The team is also exploring how to make this kind of distributed decision-making available through simpler APIs, see Sakana AI's Fugu router that handles model selection automatically.

The team notes that the approach is modular by nature. The same infrastructure and identical per-cube code that handles shape classification also extends to localizing missing modules and guiding regrowth. Scaling appears to be predictable: as spatial resolution and structural complexity increase, maintaining high performance requires greater model capacity, but the relationship is favorable.

The work is, for Sakana AI, a first step in a broader direction: taking the principles of collective intelligence studied in software and letting them emerge, decentralized and robust, in the physical world. The paper in Nature Communications is open access, and the code is available on GitHub. The team also flags limitations, including the fact that narrow structural bottlenecks (a single failure along a guitar neck) can break the classification process more easily than in shapes with redundant connectivity.