基于视觉的抓取论文分享Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp Synthesis Approach

Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp Synthesis Approach

Abstact

1. real-time, object-independent grasp synthesis method used for close-loop grasping
2. one-to-one mapping from a depth image (2015 IJRR的文章有做过实验，发现depth image对抓取的预测至关重要) overcomes limitations of current deep-learning grasping techniques by avoiding discrete sampling of grasp candidates and long computation times(选取candidate的方法非常的耗时，本文直接去掉了这个方式)
3. light-weight && single pass ( 50hz for close-loop grasping)
4. enabling accurate grasping in non-static environments where objects move and in the presence of robot control inaccuracies
5. real world test (unseen object 83%, moving household object 88%, dynamic clutter 81%)

Introduction

1. in order to perform grasping and manipulation tasks in the unstructured environment && dynamic environment of real world, a robot must be able to compute grasps for almost unlimited number of objects.
2. deep-learning takes biggest advancement in grasp synthesis for unknow items.
3. most methods using deep-learning technique are based on adapted versions of CNN architecture designed for Object detection, it computation times(1s - 10s) and rely on precise camera calibration and precise robot control, even in static environment
4. GG-CNN, use “generative” to differential our direct grasp generation method from methods which sample candidates.
5. GG-CNN has twofold advantages over SOTA CNN based grasp synthesis:
   • directly generate grasp poses on pixelwise bases, use semantic segmentation rather sliding windows or bounding boxes
   • has orders of magnitude fewer parameters than SOTA, 19ms in PC with GPU, fast enough for close-loop grasping
6. close-loop methods has obviously advantages

Related Work

1. Grasping Unkown Object
2. Closed-loop Grasping
3. Benchmarking for Robotic Grasping

Grasp Point definition

consider detecting and executing antipodal grasps on unknown objects, perpendicular to planar surface.

Experimental Set-up

• Physical Components

  • Kinova 6DOF robot fitted with Kinova KG2 2-fingered gripper
  • Realsense SR300 RGB-D camera(80mm above the closed fingertips and inclined at 14° towards the gripper)
  • PC with Ubuntu 16.04 with 3.6GHz i7-7700 Nvidia GeForce GTX 1070(6ms single depth image, 19ms entire grasping pipeline)

  (Realsense has a specified minimum range of 200 mm, in reality > 150mm, because camera’s infra-red projector and camera cause shadowing in the depth image caused by the object, Kinova KG-2 gripper has maximum stroke of 175mm, minimum of 15mm)

• Test object

  • Adversarial Set
  • Household Set

• Grasp Detection Pipeline

  • image process: crop to square, and scaled to 300 * 300 pixel, inpatient invalid depth values using OpenCV
  • evaluation of the GG-CNN (produce the grasp map and filter grasp map with Gaussian Kernel)
  • computation of grasp pose(use both open loop and close loop. in close loop evaluation, in order to avoid rapidly switching between multiple similarly-rank good quality grasp, select 3 grasps from the highest local maxima of grasp map and select the one which is closest to the grasp usedon the previous iteration)

Experiments

1. Static Grasping
2. Dynamic Grasping
3. Dynamic Grasping in Clutter
4. Dynamic Clutter Objects

Conclusion

Our System is able to gain SOTA results in grasping unknown, dynamic objects, including objects in dynamic clutter.

Closed-loop grasping method significantly outperforms an open-loop method in the presence of simulated robot control error

最后

以上就是野性金针菇为你收集整理的基于视觉的抓取论文分享Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp Synthesis Approach的全部内容，希望文章能够帮你解决基于视觉的抓取论文分享Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp Synthesis Approach所遇到的程序开发问题。

如果觉得靠谱客网站的内容还不错，欢迎将靠谱客网站推荐给程序员好友。