Neural networks enable modular robots to reconfigure using only local neighbor data

A team of researchers has shown that swarms of pivoting cube-shaped modular robots can successfully reconfigure themselves into desired shapes by using only locally available information—each cube "sees" only its immediate neighbors rather than the entire ensemble.

The Challenge of Local Coordination#

Modular robot systems face a fundamental tension: coordinating a large group to move together efficiently typically requires each unit to know about the global state of the entire system. However, in practice—especially for systems deployed in space or other challenging environments—acquiring full-system information at every timestep is expensive and often impractical. This research asks whether local awareness alone could be sufficient.

How the Approach Works#

The researchers trained neural networks using reinforcement learning to control individual cubes in a two-dimensional setting. While decisions about which cube should move next remain globally coordinated, each cube's controller operates using only sensory input from cubes in its immediate neighborhood. The neural networks learn to accumulate information across multiple message-passing rounds between neighbors, allowing cubes to indirectly gain insight into the broader ensemble configuration.

Architecture and Symmetry#

The team also explored whether embedding geometric symmetries—rotational and mirror symmetries of the grid pattern—directly into the neural network architecture would improve performance. These symmetry-aware designs did offer modest benefits during training and enabled smaller, more efficient models, though the gains were limited.

Key Findings#

Even the most locally constrained versions of the system successfully reached target configurations. Reconfiguration speed scaled predictably with the amount of information available to individual cubes: nearest-neighbor-only systems worked but took longer, while arrangements with broader neighborhood awareness converged faster. Crucially, near-optimal reconfiguration could be achieved with only nearest-neighbor interactions if the network performed multiple rounds of information exchange between cubes.

Broader Applications#

The researchers note their method transfers to other systems relevant for space applications, including sliding-cube modular robots and autonomous satellite swarms (CubeSats). The approach demonstrates that local-only sensing and coordination mechanisms can enable effective large-scale reconfiguration, potentially opening new possibilities for autonomous systems in environments where global communication is limited.