GraspVDN: scene-oriented grasp estimation by learning vector representations of grasps

Abstract Grasp estimation is a fundamental technique crucial for robot manipulation tasks. In this work, we present scene-oriented grasp scheme taking constraints of the pose imposed by environment into consideration and training on samples satisfying constraints. We formulate valid grasps parallel-jaw gripper as vectors in two-dimensional (2D) image detect them with fully convolutional network that simultaneously estimates vectors’ origins directions. The detected are then converted to 6 degree-of-freedom (6-DOF) tailored strategy. As such, able multiple candidates from cluttered scene one shot using only an RGB input. evaluate our approach GraspNet-1Billion dataset archived comparable performance state-of-the-art while being efficient runtime.