Universal Robots Robotic Arm – What technologies are involved in intelligent gripping of robotic arms?

The gripping needs of the robotic arm to confirm the position of each section of the robotic arm

First of all, the robotic arm needs a visual servo system to confirm the orientation of the object, which can be divided into two systems: Eye-to-Hand and Eye-in-Hand, according to the relative orientation of the end actuator (hand) and the vision sensor (eye).

The Eye-to-Hand is detached, and the field of view is fixed, so if the calibration accuracy of the camera is high, the accuracy of the visual positioning and grasping will also be higher.

Eye-in-Hand holds the robotic arm to the vision sensor, and the field of view changes as the arm moves, and the closer the sensor is, the higher the accuracy, but too close to the point of view may make the policy out of the field of view.

The cooperation of the fine vision system and the flexible robotic arm can complete a perfect grasp, and this is the central problem in the operation of the robot at that time, which can be summed up as one thing: find the right gripping point (or adsorption point) and grab it. After that, the transfer fulfillment is attributed to the branch of movement planning.

Now several solutions for the main stream

Model-based

This method is very easy to understand, that is, to know what to grasp, choose the method of physical scanning in advance, and give the data of the model to the robot system in advance, and the machine only needs to carry out less calculations in the actual grasping:

1. Offline accounting: according to the end type of carrying, calculate part of the grasping point for each object model;

2. Online perception: through RGB or point cloud map, calculate the three-dimensional pose of each object;

3. Accounting grasping point: under the coordinate system of the real world, according to the requirements of anti-collision, the best grasping point of each object is selected.

The RGB color space is composed of three basic natures: red, green, and blue, which are superimposed into arbitrary colors, and similarly, arbitrarily a color can also be disassembled into a combination of three fundamentals, and the robot understands the "color" through the color coordinate value. This method is similar to the direction in which the human eye recognizes colors, and is widely used on display screens.

Half-Model-based

In this practice, it is not necessary to have a thorough prediction of the object to be grasped, but many similar objects are required to practice the algorithm, so that the algorithm can usefully "cut" the drawing in the pile of objects and identify the edges of the object. This approach to practice requires these processes:

1. Offline practice of picture cutting algorithm, that is, to distinguish the pixels in the picture according to the object, this kind of work is generally handled by a special data labeler, according to the needs of engineers, marking out the different details in the massive pictures;

2. Process the drawing cutting online, and find a suitable grasping point on the artificially marked object.

This is a method that is now more widely used, and it is also the main thrust that the robotic arm grasps to propulsion. The development of robotic arm skills is slow, but the drawing cutting of the accounting machine vision is progressing quickly, and it has also leveraged the development of robots, unmanned driving and other occupations from the side.

Model-free

This practice method does not involve the concept of "object", the machine directly calculates the appropriate grasping point from the RGB drawing or point cloud map, the basic idea is to find the Antipodal (contrast point) on the picture, that is, there is a point that may "catch up", and gradually practice the grasping strategy. This practice method often allows the robot to try a lot of different kinds of objects and carry out self-supervisedlearning, and Google's Arm Farm is one of the representatives.

There are two types of dividends according to the crawling method: according to the analytic method and according to the data-driven method.

Analysis method: Mainly based on the analysis of dynamics and geometry, it is generally required to know the model of the object (known object), which is summarized as Professor Li Zexiang's "Introduction to Mathematics of Robot Operation". Based on the spinor theory, the book introduces many of the basics of robot manipulation. The profiling method helps us understand the entire gripping process and what physical quantities affect the stability of the gripping.

Data-driven: Mainly based on deep learning and reinforcement learning methods. Allows the robotic arm to grab objects that do not know the model (Unknown object) and objects that it knows (familiar object). Therefore, it is based on the development of deep learning, and it is still a relatively new field of research, and no relevant works have been found so far (if you have expectations, you can recommend it). The more classic introductory series is: the field of study, and no relevant works have been found so far (if you have expectations, you can refer to it). The more classic starter series is: the Dex-net series. Code, datasets, and engineering videos are provided, and the paper is clear. (Dex-Net by BerkeleyAutomation) Data-driven can be divided into three approaches: learning from human demonstration, learning from labeled training data, trail and error. The skills involved are: imitation learning, deep learning, reinforcement learning.

At the same time, according to the characteristic classification of the grabbed object, we can divide into three categories: know object, familiar object, and unknown object.

The general pipeline for known objects is: