深層強化学習によるタコ型レスキューロボットの平面直線移動

抄録

Octopus rescue robot is proposed as a new rescue robot that can move and work in confined spaces. The robot has eight legs, each with ten joints, and a combination of alternating roll and yaw angles. Since there were no studies analyzing octopus movement on land, we wanted to enable linear movement by using deep reinforcement learning in this study. By giving each leg an motion of the sine wave and learning the amplitude and phase of the sine wave, the number of outputs actions was reduced to 16 and the learning was converged. Deep reinforcement learning was performed with a maximum step count of 12M steps, which enabled the robot to move in a straight line at an average speed of 0.063 m/s with almost no displacement.