Online Fast Reachable Grasping Pose Estimation in Changing and Constrained Environments Based on Active Learning

In response to the low grasping success rate of robots in changing and constrained environments, a fast reachable pose estimation algorithm (FRPEA) is proposed for estimating the reachable grasping poses. This algorithm uses support vector machine and active learning to rapidly classify and estimate reachable poses online. It provides a reachable and collision-free grasping pose before motion planning, which avoids motion planning failures caused by unreachable grasping poses or collisions. Aiming at the similarity of adjacent objects' grasping poses, FRPEA uses a designed sample fast update strategy rather than resampling based on the support points of the original object to generate a small number of new samples quickly. It addresses the issue of slow online computation caused by extensive inverse kinematics calculations and collision detection. FRPEA selects the reachable pose that minimizes the rotation of the manipulator's joints to grasp objects. The sampling, labeling, and training are all completed online to adapt to changing and constrained environments. Multiple experiments are conducted in simulation, laboratory, and plant factory. The results show that in changing and constrained environments, FRPEA is ten times faster than existing online grasping pose estimation algorithms, and the grasping success rate is significantly improved compared to the existing algorithms based on offline training.