Classification and discussion of plane grab detection methods based on learning

The task of plane grab detection is , Input perceptual data , Output crawl configuration . So far , There are mainly two kinds of learning based plane grab detection methods ：

（1） A one-stage end-to-end learning approach .

（2） Two stage learning method .

1、 One stage learning

In this kind of method , Directly learn the mapping function from input data to crawl configuration , Neural networks . The input is usually RGB Image or depth image , Output is Crawl the set of configurations , Then select the optimal crawl configuration according to the confidence . Currently based on Rectangle grabbing indicates The main method is , Such as

（1）Fully Convolutional Grasp Detection Network with Oriented Anchor Box

（2）Real-world Multi-object, Multi-grasp Detection

（3）Densely Supervised Grasp Detector (DSGD)

2、 Two stage learning

It includes two stages ： Grab configuration sampling , Crawl configuration evaluation . In the first stage, multiple crawl configurations are sampled , The second stage evaluates the quality of each candidate crawl , Then choose the best grab . At present, the best is Dex-Net 4.0：

（1）Dex-Net 4.0: Learning ambidextrous robot grasping policies

3、 Discuss

One stage learning Of GroundTruth It is all the feasible grab configurations on the object , Annotation grab configuration can be manually annotated 、 Robot actual trial and error annotation or automatic generation of virtual environment , There are several problems ：

（1） There is no unified standard for manual annotation , There is an error .

（2） Robot tagging is too time-consuming ,2016 year ICRA Of Supersizing Self-supervision: Learning to Grasp from 50K Tries and 700 Robot Hours It took a lot of time , The final result is unsatisfactory .

（3） No matter which annotation method , It is impossible to mark all the feasible grasping methods of the object in detail , It means , Neural networks do not have the best learning target.

（4） Multi object stacking scenes are difficult to label , As a result, the current plane grab data sets are almost all single objects (cornell、Jacquard etc. , Now there should be a synthetic multi-object scene data set , Not paying much attention ), Make the learning network in bin-picking Poor performance in the task .

Two stage learning Of GroundTruth Is the quality of the grab sample , namely 0 or 1. Datasets can automatically generate datasets in a virtual environment , Study of the target And the best , But synthetic image Dex-Net4.0 That kind of millions of samples is not what ordinary laboratories can do . For all that , It can be seen that the two-stage learning method is better than the one-stage learning method , More worthy of study . The next research direction is how to perform crawl detection for small data sets .