ベイズ最適化を用いた移動マニピュレータによるドア開け動作の生成

抄録

We propose a new method for solving the learning problem in the door opening task based on Bayesian optimization. The method is evaluated in a simulation of a door with rotating handle and an omnidirectional mobile robot with two 6-DoF manipulators. In this method, the objective function of Bayesian optimization runs a simulation where the robot attempts to open the door by moving the grip in a direction indicated by a five-dimension problem space. The value of the objective function is the time until the robot encounters forces or torques above a certain threshold. The results indicate that learning the sequence of movements necessary to open a door requires no more than 15 trials for each movement.