Research‎ > ‎Robot Arms‎ > ‎Robotic Application‎ > ‎


Robot task using artificial intelligence


- Research Background

Logistics systems have many kinds of goods so that workers work around the warehouse to pack and deliver products. This consumes a lot of time and manpower.


- Research Objectives

Robots deal with various types of objects using artificial intelligence techniques.


- Research Output

 Object detection based on the Faster R-CNN (22 types of objects, mAP: 83.5%)

Object grasping using robot arm based on the AlexNet (Grasping success rate: 79%)


- Object detection

* Need to know type and position of objects in a short time using a RGB camera.

* Faster R-CNN

- Using VGG-Net and RPN (Region Proposal Network)

- VGG-Net is the object classification algorithm.

- Sharing the CNN architecture to extract the feature map

- RPN (Region Proposal Network) to find the bounding boxes of objects

* Training performance

- 22 types: pen, bottle, gear, glue, piston-shaft, cell phone, etc.

  - mAP: 83.5%

- Object Grasping

* Learning gripper pose of robot according to object pose


* Collecting training data in a real working environment

- It takes a lot of time.

  - Robot operation cost incurs.

  - Robot needs continuous user interventions.


* Implementing a realistic working environment on the simulator

  - Collecting a large amount of data through simulation

* Learning to grasp based on AlexNet

  - Input: image of the detected object

  - Output: pose of gripper with the highest grasping success rate

* The ensemble learning method combines various classifiers to achieve better performance than a single classifier.

- Combining the results of learning models

- Obtaining more reliable results than a learning model

- Diversity among learning models is required.

- 4 classifiers based on four data sets with ensemble learning