Sakana AI, the Tokyo lab built around evolutionary and swarm approaches to AI, has taken its collective-intelligence research off the screen and into hardware. In research published July 13 in Nature Communications, the lab and collaborators at the IT University of Copenhagen and Autodesk assembled roughly 200 physical bricks into chairs, tables, cars and boats, then had the structure figure out what shape it had become using only messages passed between physically touching bricks. No brick knew its own position, and no central processor was involved.

The result speaks to a real constraint in robotics and modular hardware. Centralized controllers are single points of failure, and wiring a position sensor into every unit of a large modular structure gets expensive fast. A system that classifies its own shape and flags its own damage using only peer-to-peer signals is a candidate architecture for warehouse racking, deployable shelters, or spacecraft components that need to keep functioning after one piece breaks or drops offline.

Each brick runs an identical local model called a neural cellular automaton, a network in which every cell updates its internal state from its immediate neighbors alone, then repeats the process over many cycles until the whole structure settles on an answer. The bricks use 3D convolutional layers to read neighbor states and update them iteratively. Rather than memorizing exact assemblies, the network learns categories: cars, chairs, tables, boats, planes, houses and guitars.

Hardware trials assembled four shapes ranging from 26 to 197 units and hit a 100 percent classification success rate across all of them. Convergence took under 60 update cycles, roughly three minutes in real time. Simulation runs across a broader set of shapes reached 98.97 percent accuracy, and the physical bricks performed comparably despite the noise and imperfect contact that hardware introduces.

The system tolerated damage unevenly. Most shapes held their accuracy at a 5 percent brick-failure rate. The plane and the boat degraded only slightly even at a 15 percent failure rate, a level that would cripple most sensor networks. The guitar did not fare as well: its narrow neck concentrated so much of the shape’s connectivity into a handful of bricks that losing any single one of them broke classification.

That bottleneck failure is the paper’s most useful caveat. Real deployments rarely offer the clean, evenly distributed connectivity of a chair or a plane, and pinch points like the guitar’s neck may be the norm in irregular structures rather than the exception. Anyone porting this approach to an actual product should stress-test connectivity chokepoints before trusting the failure-tolerance numbers.

The network generalized past its training set, correctly classifying a five-legged table and a boat with a shifted bridge that it had never seen before. A joint classification-and-damage task reached 98.9 percent shape accuracy while flagging damage with 94.8 percent accuracy, and a growth algorithm let the system rebuild its classification of a full shape starting from a small seed cluster of bricks, a configuration it was never trained on.

Sakana AI has not disclosed a cost per brick or a timeline for moving past laboratory shapes into irregular, real-world structures, and the tested configurations topped out at under 200 units, several orders below what a warehouse rack or a building facade would require. Teams evaluating decentralized robotics or self-monitoring modular hardware should read this as an early proof of concept for the coordination method itself, not yet as evidence that it scales to production-sized structures.

Sakana AI published this research on its blog on July 13, 2026, alongside the underlying paper in Nature Communications.