Smart Cellular Bricks: Towards Collective Intelligence for the Physical World

Sakana AI, along with researchers from the IT University of Copenhagen and Autodesk, just published work in Nature Communications describing a swarm of tiny cube-shaped circuit boards that behave less like a robot and more like a tissue. Each brick is identical, runs the same small neural network, and only exchanges signals with whichever bricks it happens to be physically stacked against. No brick knows where it sits in the overall structure or what shape it's part of. And yet, cycle after cycle, the whole assembly converges on an answer: this is a table, that's a guitar, this one's a boat.

The underlying method is called a Neural Cellular Automaton, a learned version of the old cellular automata idea where simple local rules produce complex global behavior. Instead of hand-coding those rules, the team trained a neural network to update each cell's internal state based on nothing but its own memory and whatever its immediate neighbors are broadcasting. In simulation, with more than 500 virtual bricks, the system hit 98.97% accuracy classifying object types like planes, chairs, cars, houses and guitars. Moving that same trained model onto real hardware, with a guitar built from 26 bricks and a round table from 197, the bricks reached a correct consensus on all four physical shapes tested, a perfect record, and did it in under 60 update cycles, roughly three minutes.

What makes this more than a neat parlor trick is how it holds up when things break. The researchers disabled subsets of bricks so they couldn't send or receive messages, then watched what happened to classification accuracy. Most shapes barely flinched at a 5% failure rate, and the plane and boat stayed accurate even at 15%. The guitar was the weak link: because its neck is a narrow structural bottleneck, killing a single brick there could sever communication between the two halves and throw off the whole classification. It's a reminder that decentralized robustness isn't automatic everywhere, geometry still matters.

The system also generalized to shapes it had never explicitly seen, a table with five shortened legs in random spots, a boat with an off-center bridge, both classified correctly. A scaled-down table, though, got mistaken for a chair, likely because shrinking the brick count compressed away the structural cues the network relies on. Looking inside the hidden channels, the researchers found something that echoes actual biology: early on, the bricks establish left-right and radial gradient patterns reminiscent of morphogens guiding embryonic development, and for chairs specifically, a signal spreading outward from the backrest is what eventually tips the group away from a default "table" guess.

Beyond spotting damage, the bricks were trained to locate it and, in a further step, to regrow around it. Starting from just a small seed cluster, cells kept adding new modules in whatever direction their neighbors pointed until no more damage was flagged, recovering shapes with high accuracy even though the model was never trained to start from a handful of cells. Joint training for classification and damage detection barely dented accuracy, 98.9% on shape classification versus 94.8% on spotting damage, and simulations pushed the idea further still, scaling up to 64x64x64 grids and assemblies of more than 18,000 cubes with shapes that included internal cavities and hollow structures.