<b>graspness measure<br></b><br>point-wise graspness scores & view-wise graspness scores
Single Object Graspness
采样点选取<br>
夹爪深度<br>
夹爪旋转<br>
force analytic mode
Scene-Level Graspness
问题1：抓取位姿与背景物或者临近物体干涉<br>
GraspNet-1Billion: A Large-Scale Benchmark for General Object Grasping
reconstruct the scene using object 3D models and correspond-<br>ing 6D poses
碰撞检测
公式<br>
Ckij-collision label<br>
问题2：RGBD观测到的是单视野局部点云
associate the scene point cloud with the projected object point
通过对象的6D Pose将对象点投影到场景中<br>
<i>cascaded graspness model<br></i><br>作用：提取每个点的局部特征向量<br><br>局部特征向量：<br>
Two sub-function
F - 高维特征
point level过滤掉大部分不可抓取点；View-wise只需要计算剩下的点<br>
backbone network<br>
<b>ResUNet14</b><br>extraction of both global and<br>local point features
功能<br>
Graspable Farthest Point Sampling
功能
<b>使用MLP建模</b><br>to generate point-wise graspable landscape.
select points with graspness score larger than δp
adopt farthest point sampling (FPS) to maximize distances among<br>sampled points
Graspable Probabilistic View Selection
功能
<b>使用MLP建模<br></b>to the sampled seed points and output M × V vectors for view-wise graspable landscapes<br>and M × C residual features for grasp generation.
<b>Fibonacci lattices生成V个观测视野</b>
grasp operation model
<b>Crop-and-refine</b><br>从点云中裁剪出潜在的抓取区域，然后对这些区域进行细化处理，以精确估计抓取位姿和提高抓取检测的准确性。通过这种方式，可以从复杂的场景中有效地识别和定位可抓取的物体。
Cylinder-Grouping from Seed Points
Grasp Generation from Candidates
<b>PointNet</b><br>for grasp generation
Grasp Score Representation
minimum friction coefficient
Loss Function<br>
Loss Function说明<br>
