Transferable Active Grasping and Real Embodied Dataset

Robotics Dexterous Grasping: The Methods Based on Point Cloud and Deep Learning

“…By adopting sampling method in GPD proposed by Pas et al (2017), Lopes et al (2018), Schnaubelt et al (2019), Bui et al (2020), Chen et al (2020), and Deng et al (2020) sample the grasp points in point cloud for candidates generation. Lopes et al (2018) find the largest planar surfaces which is infeasible for grasping by using RANSAC (Fischler and Bolles, 1981) and isolates the closest object to the camera from the rest of the scene to obtain object segmentation based on min-cut (Golovinskiy and Funkhouser, 2009).…”

“…Deng et al ( 2020 ) detects and segments the object from RGB-D images based on PoseCNN (Xiang et al, 2017 ), then a sampling method in Eppner et al ( 2019 ) is adopted to generate 100 candidates for assessment and execution. Chen et al ( 2020 ) utilizes object segmentation for mask-guide to improve the precision of sampling. Lin and Cong ( 2019 ), Lin et al ( 2019 ), Sun and Lin ( 2020 ), and Yu S. et al ( 2020 ) follow the same idea in GPD with additional physical or geometric constraints.…”

“…At each time stamp, it generate several grasp candidates based on hierarchical sampling and then one pose will be chosen by the learned policy. Chen et al ( 2020 ) adopts RL to obtain an appropriate viewpoint based on the mask-guided award to perform GPD module. Rather than deploying RL strategy on grasp planner, this model focuses on acquire a better view sight to improve the grasp accuracy.…”

“…The methods mentioned in sections Grasping Candidate Generation and Grasp Candidate Evaluation based on CNN to generate and evaluate grasp candidates show the potentials of generalization (Mahler et al, 2017 , 2018 ; Yan et al, 2019 ; Chen et al, 2020 ; Shao et al, 2020 ). However, for the proposed approaches based on reinforcement learning are hard to be transferred to other tasks.…”

Learning Pick to Place Objects using Self-supervised Learning with Minimal Training Resources

“…As a result, an end-to-end 6-DoF closed-loop grasping model using RL is shown employing a learned value function (Q-value). Also, an RL framework and 3D vision architectures were proposed [27] using handmounted RGB-D cameras. However, manipulation with more task-dependent representations must be learned from limited training data.…”