環境変化に臨機応変に対応する協調搬送行動のためのAttention-based Neural Networkの設計

抄録

In order to design a controller for cooperative transportation by swarm robots, there is an approach to optimize a Neural Network (NN) as a controller using reinforcement learning. In most of the previous studies, the NN used the full coupling layer, which cannot cope with changes in the number of robots and states observed by the robots. In this study, we adopt a NN with the Self-Attention mechanism proposed in machine translation as the controller of the robot. Since this mechanism can respond to changes in the number of inputs, it has the potential to cope with environments where the amount of state observed by the robot is different from that during training. In this paper, we train the cooperative transport system in an environment with three robots and three obstacles, and verify whether the system can respond to changes in the number of robots and obstacles using the controller.