Reinforcement-Learning-Based Task Planning for Self-Reconfiguration of Cellular Satellites

Cellular satellites, which are composed of many standard unit cells, represent a class of novel and promising satellites for future space explorations. Their potentials have been well recognized in the aerospace field. The most attractive feature of cellular satellites is the on-orbit self-reconfiguration capability through cell-by-cell moves. However, it is extremely challenging for a cellular satellite to autonomously achieve the optimal self-reconfiguration with fewest cell moves, because the search space for legal actions may be larger than that of the game of Go if the satellite has a certain number of cells. In this article, we propose a reinforcement learning-based task planning strategy for the self-reconfiguration of cellular satellites. Inspired by the recent progress of AlphaGo and AlphaGo Zero, we calculate the cell move sequence and predict the cell placements in the self-reconfiguration process by combining the Monte Carlo tree search and the neural network. The reinforcement learning-based task planning strategy is validated by comparing with the traditional melt-sort-grow algorithm. The validation results demonstrate that the proposed strategy can significantly reduce the number of cell moves for the self-reconfiguration of cellular satellites.

Automated summary

Cellular satellites, composed of standard unit cells, have the potential to be a promising class of satellites for future space explorations. This article proposes a reinforcement learning-based task planning strategy for the self-reconfiguration of cellular satellites, inspired by the progress of AlphaGo and AlphaGo Zero. The strategy combines Monte Carlo tree search and neural network to calculate cell move sequence and predict cell placements, resulting in significantly reduced number of cell moves.