Cubic bricks self-classify shape via neural cellular automata

A team including Sebastian Risi has built physical modular robots that figure out their own shape by talking only to their neighbours. The system, described in a paper on arXiv and in Nature Communications, uses simple cubic bricks that each run the same small neural network. No central controller, no global map, no per-brick position information. From local messaging alone, the collective decides whether it is a plane, a guitar, a boat or a table.

The mechanism is Neural Cellular Automata. Each cube executes the same policy and, iterating with its immediate neighbours, converges on a shape label in fewer than 60 update cycles, roughly three minutes on hardware. In simulation the authors report 98.97% classification accuracy across the four shape classes. The physical prototypes, ranging from 26 bricks for the guitar to 197 for the round table, correctly identified all four shapes.

The interesting bit is what happens when the swarm is degraded. At 5% module failure most shapes still classify correctly, and for the plane and boat the reporting describes only minimal degradation even at 15% failure. The guitar is the honest exception; its thin neck makes localised faults there disproportionately damaging. Extending the same framework to damage detection gave over 90% accuracy, and the recovery routine has surviving cells predict one of six axis directions (-X, +X, -Y, +Y, -Z, +Z) to grow new modules into, repeating until the shape is restored.

Take the specifics as reported rather than settled. The physical scale is still modest at close to 200 cubes, well below the 500+ used in simulation, and the retrieved material does not go deep on per-brick cost, power draw, or communication overhead at scale. Nor is it clear how the system generalises to shapes far outside the four demonstrated classes.

The reason to watch it is the direction of travel. If distributed neural policies can turn dumb structural units into a swarm that recognises and heals itself, the eventual payoff sits in things like on-orbit assembly, self-diagnosing infrastructure and reconfigurable manufacturing, where the alternative today is either central control that fails hard or handcrafted rules that do not survive contact with the real world. The authors' own framing is that this "mirrors the decentralised strategies seen in morphogenetic processes," and that is the framing worth remembering as the brick count grows.