Smart Cellular Bricks: Towards Collective Intelligence for the Physical World

Collections of hundreds of physical 3D cellular bricks, simple modular hardware units that run identical local Neural Cellular Automata without any global knowledge, collaboratively infer their overall shape class.

（＊日本語は英文の後に）

At Sakana AI, a recurring theme in our research is collective intelligence: the observation that sophisticated, robust behavior can arise from many simple parts following local rules, with no central controller, as it does in a colony, a tissue, or a brain. Until now, we have explored this idea in simulated systems, such as getting several frontier models to reason together and build on one another’s attempts, coordinating them so that many models act as one, or having agents with partial, overlapping views negotiate along their shared boundaries to converge on a globally consistent solution.

Today, we are happy to share that a paper extending this line of work into physical hardware has been accepted for publication in Nature Communications. The system is a collection of simple cubic bricks, each running the same small neural network and communicating only with the bricks it is physically connected to. No brick is told its position or which shape it is part of, yet from these purely local exchanges the collective converges on the correct global shape and can locate where modules are missing or damaged.

For us, this is a first step toward taking our work on collective intelligence beyond software and into the physical world. We wanted to ask whether the same decentralized principles hold up when communication is noisy and modules fail. They largely do. Hundreds of bricks classify a range of 3D shapes, still recognize shapes with variations they were not trained on, and keep converging correctly even when a fraction of their modules go silent. Using the same framework, the collective can also flag structural damage and guide a step-by-step recovery.

The work is a collaboration between researchers at IT University of Copenhagen, Sakana AI, and Autodesk.

Read the full paper in Nature Communications: https://www.doi.org/10.1038/s41467-026-75166-7

Explore the code on GitHub: https://github.com/rmorenoga/cube3D

Introduction

Many biological systems exhibit a remarkable capacity to determine their own anatomical structure. Through local communication and self-organization, groups of cells can assess whether they have correctly formed a target shape, such as an organ, and can actively remodel body parts following injury. A salamander can regenerate a damaged tail that transforms into a functional leg, and simple organisms like Hydra and Planaria can fully restore their morphology regardless of which part is lost. This ability to classify general anatomical features, rather than match a fixed target shape, is what enables variability among individuals while enhancing the robustness of the process. The overall function and design of an organ may be consistent across a species, even as its specific shape, size, or scale differs from one individual to the next.

Artificial systems composed of many physically distributed modules that can autonomously infer their structural class, without centralized control, would represent a significant step toward more adaptable, intelligent artificial collectives. Such systems could enable powerful applications in smart materials and reconfigurable robotics, where global knowledge must emerge from local sensing and communication. Motivated by the scalability and resilience of biological collective intelligence, we introduce a fully decentralized system in which hundreds of physically embodied “cellular” bricks collectively classify their global shape and detect local damage, with no central controller and no module ever knowing its own position.

Neural cellular automata for shape classification. (A) A cellular brick module. (B) Bricks assembled into four different shapes. (C) Each cell takes in local information from its connecting neighbors and its own hidden channels; information is aggregated locally, enabling the object to recognize its particular shape over multiple iterations. (D) The local update rules are encoded with a neural cellular automaton, a deep neural network.

The collective intelligence algorithm we developed builds on the framework of differentiable Neural Cellular Automata (NCA) and self-classifying collective systems, extended to operate in 3D and implemented on physical modular hardware. NCAs generalize traditional cellular automata, in which the local update rules are typically hand-crafted, by instead learning these rules. Unlike traditional CAs that operate with discrete cell states, NCAs use continuous-valued cell states, enabling end-to-end differentiability and compatibility with gradient descent. On a high level, each cell in our system is tasked with determining which type of shape it is a part of, based solely on communication with its local neighbors and its own memory state. The update rules are parameterized by a deep neural network built from 3D convolutional layers, whose outputs are added to the state of each cell. Here, the collective is tasked to distinguish between objects resembling planes, chairs, cars, tables, houses, guitars, and boats, trained with a cross-entropy loss to predict the class label through gradient-based optimization.

Rather than matching against a single, predefined configuration, our system generalizes across entire classes of shapes, such as different planes or different tables. This shift from precise self-recognition to high-level shape classification enables greater flexibility and tolerance to variation. The system can also detect structural inconsistencies caused by missing or faulty modules, using only local interactions and without requiring any actuation or centralized sensing.

Results

Each cellular brick is a small printed-circuit-board cube with electrical connectors on all six faces, a microcontroller, an LED to display its current class guess, and the electronics to power itself. Bricks stack together into arbitrary objects and communicate purely locally over a custom digital serial protocol, iterating until the collective converges on a single shape label.

We ran large-scale experiments with more than 500 bricks in simulation and almost 200 physical bricks. In simulation the approach reaches 98.97% accuracy. Transferring the simulation-trained NCA directly onto hardware, we built four distinct shapes ranging from 26 bricks (a guitar) to 197 (a round table); the bricks reached correct consensus on a plane, a guitar, a boat, and a table, with 100% success rate across all four. In hardware, the collective converges in fewer than 60 update cycles, about three minutes of real time.

Robustness to faulty cells and shape variation

Biological systems are remarkably robust to damage, noise, and faulty components, thanks to their local, distributed decision-making. To test whether our system inherits this property, we disabled subsets of bricks in hardware, preventing them from sending or receiving messages, and measured the effect on recognition. Most shapes maintained high accuracy at 5% failure rates, and some, like the plane and boat, degraded only minimally even at 15%. The exception is shapes with narrow structural bottlenecks: in the guitar, a single failure along the neck can sever the two halves of the object and disrupt classification.

For the guitar and plane assembled in hardware, we disabled particular cells (marked in red) to test robustness. By design, the plane is far more robust than the guitar, where a single failure along the neck can break the classification process.

The system also generalizes beyond the specific examples it saw during training. We designed test shapes with novel variations within known classes: a table modified to have five shortened legs at random positions, a boat with its central bridge shifted off-center, and scaled-down versions of the plane and table. The altered five-legged table was still correctly classified as a table, and the shifted boat bridge did not impair recognition, suggesting the NCA captures abstract structural features rather than overfitting to specific instances. It is not infallible, however: the scaled-down table was misclassified as a chair, likely because the reduced module count compresses the structural cues the network relies on.

Emerged communication strategies

We also looked at how the collective actually solves the task. Inspired by morphogens (the diffusible molecules that form gradients across developing tissues and give cells positional information), we examined the activation patterns of the NCA’s hidden channels. Early in the process, the system often establishes left-right and radial patternings that resemble the developmental axes of embryos.

The hidden channels learn morphogen-like activation patterns, such as left-right and radial gradients, enabling decentralized shape recognition and differentiation.

This also explains how the collective tells tables from chairs. Many cells in a chair are initially classified as a plane, just as in tables. Unlike tables, though, an anterior-posterior patterning is established (akin to the biological head-to-tail axis), and over time this signal propagates from the backrest outward, guiding the cells to a consensus that they form a chair. The default reading for both is “table,” and morphogen-like signals originating in the backrest gradually induce a coherent reclassification.

Damage detection and recovery

An exciting extension is to let each cell not only recognize which larger shape it belongs to, but also detect whether one of its neighboring cells has been damaged. To achieve this, the cells were trained jointly for shape classification and local damage detection, predicting either no damage or damage in one of six spatial directions. Training used synthetically damaged shapes with spherical and cubic lesions. Despite the added task, the system retained 98.9% shape-classification accuracy while detecting damage with an average accuracy of 94.8%.

Can we exploit this to recover from damage, rather than merely predict it? Biological organisms like planarians and axolotls regrow complex body parts through distributed, decentralized cycles of sensing and regeneration, with no central controller. Mirroring this, we started from a small cluster of cells and repeatedly added new cells in the direction indicated by the existing ones, until no more damage was detected. Without ever being trained starting from only a few cells, the model recovered almost all shapes across all object classes with high accuracy. Models with larger hidden states performed significantly better, likely because they can carry more information across the growth process.

Starting from a small seed cluster, cells repeatedly add new modules in the direction their neighbors indicate, regrowing a chair, a table, and a plane until no further damage is detected.

Simulation scaling to larger and more complex morphologies

Hardware experiments are naturally limited by the number of available robotic bricks and the practical constraints of physical deployment. We therefore evaluated scaling behavior in simulation, where the approach already works well beyond the morphologies demonstrated in hardware, extending from 15x15x15 to 32x32x32 and 64x64x64 grids. Moreover, the higher-resolution experiments involve substantially richer geometries than those used in hardware, including shapes with hollowness, internal cavities, and assemblies of more than 18,000 cubes. These experiments reveal encouraging scaling behavior: as spatial resolution and structural complexity increase, maintaining high performance requires greater model capacity, but the relationship is predictable and favorable.

Regeneration of objects (grid up to 64³): a fish, the Sakana AI logo, and a heart. Cells locally predict the direction of missing neighbors (colors

[truncated for AI cost control]