Introduction
On July 13, 2026, Sakana AI, the IT University of Copenhagen and Autodesk Research published a paper in Nature Communications describing something unusual: a heap of roughly 200 circuit-board cubes that can work out, among themselves, what shape they have been built into. No brick knows its position. No brick knows the shape. There is no central computer. Each cube only ever exchanges messages with the cubes it is physically touching — and after about three minutes, every LED in the assembly is lit with the same, correct answer: guitar, or table, or car.
The paper, "Smart cellular bricks for decentralized shape classification and damage recovery" (Nature Communications 17:5932), is authored by Rodrigo Moreno, Andrés Faíña, Shyam Sudhakaran, Kathryn Walker and Sebastian Risi. Sakana AI's accompanying write-up is Smart Cellular Bricks. The team calls the result the first physical realization of large-scale decentralized self-recognition in three dimensions.
What the hardware actually is
Each brick is a 44×44×47 mm cube whose six faces are 1.6 mm FR4 printed circuit boards. Inside sit an ESP32-S2 microcontroller, a DC-DC converter and a WS2812B programmable RGB LED that displays the brick's current guess; every face carries a six-pin connector. Neighbouring bricks talk through those connectors over a custom digital serial protocol. No brick has a camera, a map, or any notion of its own coordinates.
The modules do have WiFi — but it is deliberately kept out of the loop. Only after the 60 classification cycles are finished do the bricks connect to a network, and then merely to upload their logs of states and timings for the researchers. The answer itself is reached over wires, one neighbour at a time.
The intelligence lives in what runs on the microcontroller — a neural cellular automaton (NCA). A cellular automaton is a grid of cells that each update their state according to a rule applied to their immediate neighbours; Conway's Game of Life is the classic example. An NCA replaces the hand-written rule with a small neural network, and because the cell states are continuous rather than discrete, the whole thing is differentiable end-to-end and can therefore be trained with gradient descent like any other model.
Every brick runs an identical copy of that trained network on its own microcontroller — edge AI in its most literal form. Each holds a 28-dimensional state vector: one alpha channel marking whether the cell is alive, 20 hidden channels it uses to communicate, and 7 channels holding its current class logits — its vote on what shape it is part of.
How it was trained
The team trained the NCA in simulation on a 487-shape subset of ShapeNet spanning seven classes: aircraft, boats, cars, guitars, houses, tables and chairs. Voxelize a shape, seed each occupied voxel with a cell, let the cells pass messages for a fixed number of steps, and apply a cross-entropy loss to the class each cell ends up voting for. Standard supervised classification — with the twist that the classifier is the structure itself.
In simulation it reached 98.97% accuracy overall, ranging from 85.04% on houses to 99.71% on chairs.
How this differs from earlier robot swarms
This is the question worth asking, because modular robots that cooperate are not new, and the paper is careful about it.
The obvious ancestor is Kilobots — Rubenstein, Cornejo and Nagpal's 1,024-robot swarm, published in Science in 2014, which assembled itself into shapes like a star and a wrench using only local communication. MIT's M-Blocks, cubes that flip and roll into formation using flywheels and magnets, are another well-known relative.
But those systems solve the reverse problem. A Kilobot is told which shape to form and follows three hand-written rules — edge-following, gradient formation, trilateration — to get there. A Smart Cellular Brick is told nothing and infers which shape it is already part of. The paper cites Kilobots explicitly, and its framing of the gap is precise: such swarms "have illustrated distributed consensus in two dimensions, but extending such ideas to three-dimensional structures has remained challenging."
So the claim to novelty rests on three specifics, not on collective behaviour in general: the rule is learned rather than hand-crafted; the system needs no actuation at all to understand its own morphology; and it runs on ~200 physical modules in three dimensions, where comparable hardware work has been small and flat.
What happened on real bricks
The team assembled four physical shapes, from a 26-brick guitar to a 197-brick round table. In every case, all bricks reached consensus on the correct shape, converging in fewer than 60 update cycles — about three minutes of wall-clock time.
Two results are more interesting than the headline accuracy:
It generalizes structurally. Presented with objects it had never seen — a table with five irregular legs, a boat with its bridge shifted off-centre — the collective still classified them correctly, which suggests it learned abstract structural features rather than memorizing training silhouettes.
It finds its own wounds. Trained additionally to flag missing neighbours, the system located damaged or absent modules with 94.8% accuracy while keeping 98.9% shape-classification accuracy. Its robustness under failure was decent but uneven: most shapes held high accuracy with 5% of modules dead, and planes and boats survived 15%.
Two ways to heal, and the trade-off between them
Recovery — regrowing the missing part of a shape — was tested in two modes, and the difference between them is the honest heart of the paper.
In self-classifying recovery, the cells simultaneously predict the shape class and the direction of the damage (one of six Cartesian directions, or none). Nothing external is needed. But the failure mode is circular: if enough of the object is missing that it now looks like a different class, the collective confidently regrows the wrong object.
In shape-conditioned recovery, the cells are handed a compact embedding of the intended target shape. That extra bit of prior knowledge lets them tell genuine damage apart from legitimate variation within a class, and it can rebuild a structure from just a few surviving seed cells — but it means someone outside the system had to know the answer.
Fully autonomous, or reliably correct: for now, this system gets one or the other.
Where it breaks
The authors are direct about the rest of the limits, and they matter more than the demo.
The bricks are externally powered and cannot move — there is no actuation, so hardware "recovery" means recomputing a correct answer, not physically rebuilding anything. There is no role specialization: every cell stays generic forever. Scale invariance fails outright, with a shrunken table classified as a chair. And thin structures are brittle in a specific, instructive way: the guitar's narrow neck is an information bottleneck, so a single failure there severs the two halves of the shape from each other and consensus collapses.
Scaling was only demonstrated in software. Simulated grids grew from 15³ to 32³ to 64³ voxels — assemblies of more than 18,000 cubes, with hollow interiors and internal cavities — holding 0.98, 0.96 and 0.96 accuracy respectively, and still trainable on a single GPU.
Why it matters
Most work on collective and embodied AI is validated in simulation, where communication is free, latency is zero and nothing breaks. Putting the same learned local rule onto 200 microcontrollers that can only whisper to their neighbours is the step that usually exposes the gap between the two — and here it did not.
The analysis of what the bricks learned to say to each other is the most suggestive finding. The hidden channels settle into morphogen-like activation patterns — left-right and radial gradients across the structure — resembling the chemical gradients biological embryos use to tell one end of a body from the other. Nobody designed those. They fell out of gradient descent on a classification loss.
The named applications — edutainment, smart modular surfaces, eventually adaptive materials — are honest and modest. The proposed next steps are where it gets ambitious: magnetic docking, miniaturized electronics, actuation for genuine self-repair, and closed-loop autonomous growth.
Conclusion
Smart Cellular Bricks has nothing to do with large language models or scaling laws. It is a demonstration that a global property — what am I part of? — can be computed from purely local exchanges on real, imperfect hardware, by a rule that was learned rather than designed. The bricks cannot yet heal themselves, power themselves, or scale past a few hundred units outside a simulator, and their most reliable recovery mode still needs to be told the answer in advance. But they answer a question the field mostly avoids asking outside a simulator, and the code and hardware designs are public.
Sources
- Smart Cellular Bricks — Sakana AI
- Smart cellular bricks for decentralized shape classification and damage recovery — Nature Communications 17:5932 (2026)
- Programmable self-assembly in a thousand-robot swarm — Science 345:795–799 (2014)
- Code and hardware designs — GitHub: rmorenoga/cube3d
- ShapeNet dataset