Sakana AI, Autodesk demo bricks that infer their own 3D shape

A modular brick that knows what shape it is part of, without being told and without a central brain, sounds like a lab demo, and mostly is. But the demo has enough concrete numbers behind it to take seriously. In work published on Sakana AI's site, researchers from IT University of Copenhagen, Sakana AI, and Autodesk describe a system of small cubic bricks that only talk to their immediate neighbors and, collectively, infer the overall 3D shape they have been assembled into. The paper landed in Nature Communications on July 13, 2026.

The mechanism is a Neural Cellular Automaton. Each brick runs the same tiny local network, exchanges a hidden channel state with the cubes touching its faces, and after fewer than 60 update cycles, about three minutes of real time, the assembly reaches consensus on its class. In simulation the researchers report 98.97% accuracy across assemblies of more than 500 bricks. In hardware they built four objects, including a guitar with 26 bricks and a round table with 197, and all four reached correct consensus, a 100% success rate at that small sample. The same substrate can flag structural damage at 94.8% average accuracy, and some shapes, like the plane and the boat, degraded only minimally even at 15% brick failure.

The reason this reads as more than a novelty is the pitch it makes to material makers. Autodesk being on the byline is the tell. The framing is that smart materials could sense and report their own damage, the sort of capability construction and infrastructure firms have wanted since long before machine learning got involved. Embedding a distributed classifier into the building blocks themselves means the diagnostic is not bolted on afterwards.

The honest caveats are the ones you would expect. Four hardware assemblies of at most a couple hundred cubes is a proof of concept, not a scaled deployment. The researchers themselves flag that in the guitar a single failure along the neck can sever the two halves of the object and disrupt classification, so uniform-random robustness numbers do not tell the whole story. The reporting does not give cost per brick, power budget, or how well the trained rule generalizes to shapes the network never saw during training. Those are the questions that decide whether this becomes a research thread or a product line.

The direction is the part worth watching. If distributed sensing you can literally stack keeps getting cheaper and more accurate, it becomes the sort of substrate that changes what instrumented means for physical infrastructure.