Neural Cellular Automata Enable Self-Discovery of Physical Configuration in Modular Robots Driven by Collective Intelligence

Abstract # ↰

Modular robots are capable of self-assembly and reconfiguration, allowing them to adapt to changing environmental conditions and perform complex tasks. However, these robots typically require controllers that are optimized for their specific shape and intended function. In the event of an assembly failure, even minor discrepancies can lead to significant problems because the controller is designed to operate in a specific scenario. To address this issue, we propose a shape-aware embodied controller that relies on self-discovery driven by Neural Cellular Automata (NCA), which enables the system to identify its own configuration, followed by the selection of the appropriate controller, from a library of pre-trained ones, and its deployment in response to the current situation. As a result, our controller is designed to be able to adapt to unforeseen errors and even damages and continue to operate effectively even in the face of unexpected circumstances.