Neural Cellular Automaton for Decentralized Inference in Distributed Manipulation Systems.

Distributed manipulation systems employ a grid of independently controlled actuators to achieve precise manipulation of objects resting on their surface. Despite the decentralized nature of the actuators, current implementations use centralized feedback mechanisms to provide information about the object’s position to the controllers. This centralized approach introduces a potential vulnerability, as a failure in the feedback system could result in the complete failure of the system. This paper proposes an approach for characterizing objects in decentralized systems. Its validation is demonstrated through the implementation of a two-dimensional simulated network of sensing agents. These agents work collaboratively to determine a global property - the geometrical center of an object - through local communication of the information at their disposal. The method uses a Neural Cellular Automaton, a multi-agent system in which the update rule of each agent is expressed as a Neural Network, and it is a function of its neighborhood’s information. Experiments show that trained on known objects, the system displayed outstanding precision in inferring their centers and demonstrated remarkable adaptability when encountering unknown objects.