3차원 공간상의 작업-동작 동시 계획을 위한 학습 기반 동작 가능성 예측

Abstract

We propose a learning-based method to predict the feasibility of motion for the task and motion planning of pick and place tasks. The proposed method comprises three steps of tool collision checking, kinematic feasibility prediction, and path-existence prediction. We trained a support vector machine for the second step and a neural network for the third step. The proposed method can predict motion feasibility in complex workspaces with three-dimensional obstacles without discretizing action parameters (e.g., grasping pose). We verified the effectiveness of the method with runtime experiments.