Deep Learning for the selection of the best modular robots self-reconfiguration algorithm

Modular Robots Self Reconfiguration (MRSR) is one of the most challenging problems in nowadays robotics field. This problem consists in the determination of how a set of identical modular robots, with local knowledge of the system and limited energy and computational capacities, can reorganize themselves into a target topology or shape. MRSR has received great attention from the research community. Therefore, a lot of centralized and decentralized algorithms were designed to answer this problem. Unfortunately, the analysis of why and when an algorithm is better than another is less studied. In this paper, we proposed a hybrid centralized/distributed modular robots reconfiguration approach. In this approach, a convolution neural network system is used to estimate the most adapted distributed reconfiguration algorithm according to the initial shape formed by the modular robots and the target shape. Two distributed algorithms are studied: C2SR and TBSR. The designed CNN model allows determining which option is the best for a given reconfiguration problem: use of the C2SR algorithm, use of the TBSR algorithm, or both algorithms are equivalent. The obtained results show that the ML tool succeeds 97.25% of the time to determine the suitable algorithm based on the initial and the final shapes. In addition, the system can be extended to any number of algorithms. Our contribution is the production of a neural network built for the selection of the best modular robots self-reconfiguration algorithm.