Smart Cellular Bricks: 3D Neural Cellular Automata for Life‑Like Modular Robots

Smart Cellular Bricks

Each brick runs an identical 3D Neural Cellular Automata (3D NCA) network, maintains a small hidden state (Hidden Channels), and exchanges information only with its six adjacent bricks. No brick has global knowledge of the overall structure and there is no central controller.

Simulation performance

In simulated shape‑classification tasks the method achieves an accuracy of 98.97% .

Hardware demonstrations

Four physical configurations—airplane, guitar, boat, and table—were built using 26 to 197 bricks. All four reach 100% consensus on the target shape and converge in fewer than 60 update cycles (≈3 minutes of real time).

Damage detection and recovery

When bricks are removed or disabled, the distributed system first detects the abnormal region through local communication, aggregates the damage information, and the remaining bricks re‑evaluate the structure to plan recovery. Damage‑localization accuracy in simulation reaches 94.8% , and the system can adapt to damage patterns not seen during training.

Robustness

Most shapes maintain high precision with up to 5% brick failure. Even at 15% failure, airplane and boat shapes degrade only slightly. Narrow‑bottleneck structures such as the guitar neck are more vulnerable; a single missing brick can split the object.

Scalability

The NCA scales from 15×15×15 grids up to 64×64×64 grids, involving more than 18 000 cubes. Experiments include hollow and internally cavitated geometries, demonstrating that the approach can handle large, complex structures.

Limitations

The current Smart Cellular Bricks require a training phase, specialized hardware, and are limited to a predefined set of shape tasks.

Source: "Smart cellular bricks for decentralized shape classification and damage recovery", Nature Communications (2026). https://www.nature.com/articles/s41467-026-75166-7

Code example

来源：ScienceAI
本文
约2000字
，建议阅读
5
分钟
让机器人学会认识自己，发现损伤并引导自我恢复。