In recent years, data-driven services have been realized in various fields, and to achieve congestion avoidance, security of confidential information, and real-time response, edge devices that collect and process data close to the field environment are advancing the application of AI based on recurrent neural networks (RNNs). However, the constraints on the available energy and computational resources hinder the application due to the large processing load. In this study, we focused on reservoir computing, a concept derived from RNNs, which has relatively low computational complexity, and attempted to apply it to robot control. We constructed an evaluation environment using a physical simulator, with the ceiling-walking control of a drone-type quadruped robot as a target task. We also introduced a power consumption of the movement as a learning criterion, and examined a method that can optimize the power consumption during movement, and verified its effectiveness by numerical simulations.