REGNet V2: End-to-End REgion-based Grasp Detection Network for Grippers of Different Sizes in Point Clouds

TL;DR

Regnet is presented, a robotic grasping system that can adapt to different parallel jaws to grasp diversified objects, significantly outperforming several state-of-the-art methods, including GPD, PointNetGPD, and S4G.

Abstract

Grasping has been a crucial but challenging problem in robotics for many years. One of the most important challenges is how to make grasping generalizable and robust to novel objects as well as grippers in unstructured environments. We present \regnet, a robotic grasping system that can adapt to different parallel jaws to grasp diversified objects. To support different grippers, \regnet embeds the gripper parameters into point clouds, based on which it predicts suitable grasp configurations. It includes three components: Score Network (SN), Grasp Region Network (GRN), and Refine Network (RN). In the first stage, SN is used to filter suitable points for grasping by grasp confidence scores. In the second stage, based on the selected points, GRN generates a set of grasp proposals. Finally, RN refines the grasp proposals for more accurate and robust predictions. We devise an analytic policy to choose the optimal grasp to be executed from the predicted grasp set. To train \regnet, we construct a large-scale grasp dataset containing collision-free grasp configurations using different parallel-jaw grippers. The experimental results demonstrate that \regnet with the analytic policy achieves the highest success rate of $74.98\%$ in real-world clutter scenes with $20$ objects, significantly outperforming several state-of-the-art methods, including GPD, PointNetGPD, and S4G. The code and dataset are available at https://github.com/zhaobinglei/REGNet-V2.

Figures (18)

• Figure 1: (a) Different suitable grasps generated for grippers with different sizes. When grasping a bottle, (b) For a gripper with a bigger width, points closer to the bottom of the bottle have a higher grasp confidence. (c) For a gripper with a smaller width, points closer to the bottleneck have the higher grasp confidence.
• Figure 2: The REGNetV2 pipeline. SN embeds the sampled gripper parameters to the observed point cloud and utilizes them to predict the point grasp confidence. Then GRN predicts grasp proposals based on predicted grasp confidence. After RN refines the proposals, we select the optimal grasp through the analytic grasp selection strategy to execute grasping.
• Figure 3: (a) The visualization of a 2-finger parallel gripper. (b) A simplified version of the gripper.
• Figure 4: (a) The definition of a grasp. (b) When a gripper closes to grasp, two contacts $c_i, c_j$ are generated. $\alpha_i, \alpha_j$ are angles between the force direction and normals $n_i, n_j$. Blue lines show friction cones (force_closure1) generated at contacts.
• Figure 5: The architecture of REGNetV2. SN takes the point cloud and the gripper parameters as input and outputs point grasp confidence. The darker point has a higher point grasp confidence. GRN selects some points with higher confidence and predicts grasp proposals based on grasp regions centered on the selected points. RN then refines grasp proposals through the gripper closing area features and grasp region features.