Mainstream intelligent grasping solutions in robot learning

One of the classic problems in learning is sorting: taking out the target item from a pile of disordered items. For a courier sorter, this is almost a thoughtless process, but for the arm, it means complex matrix calculations.

Express sorters are sorting

In fact, difficult mathematical problems that require humans to spend a lot of time are very easy to handle with the system, but the sorting action that can be performed with almost no thinking is a hot topic for robot research experts around the world.

Robotic arm grasping requires determining the position and posture of each segment of the robotic arm

First of all, the robotic arm requires a visual servo system to determine the position of the object. According to the relative position of the end effector (hand) and vision (eye), it can be divided into two systems: Eye-in-Hand.

Eye-in-Hand fixes the robotic arm and visual sensor together. The field of view changes as the robotic arm moves. The closer the sensor is, the higher the accuracy. However, if it is too close, the target may be out of field of view.

The cooperation between the precise visual system and the flexible robotic arm can achieve a perfect grasping, which is the core problem in the current robot operation. In summary, it is just one thing: find the right grasping point (or adsorption point) and grasp it. The subsequent transport execution belongs to the branch of motion planning.

Several mainstream solutions

Model-based

This method is easy to understand. It means knowing what to grasp, scanning the object in advance, and providing the model data to the robot system in advance. The machine only needs to perform fewer operations in the actual grasping:

2. Online perception: Calculate the 3D pose of each object through RGB or point cloud images;

3. Calculate the grasping point: In the real-world coordinate system, select the best grasping point for each object based on requirements such as collision avoidance.

The RGB color space is composed of three basic colors: red, green, and blue. They can be superimposed to form any color. Similarly, any color can also be a combination of the three basic colors. Robots understand "color" through color coordinate values. This method is similar to the direction in which the human eye recognizes color and is widely used on display screens.

CGrasp for random grasping of precision bearings

Half-Model-based

In this training method, there is no need to fully predict the objects to be grasped, but a large number of similar objects are needed for training, so that the algorithm can effectively "segment" the image in the pile of objects and identify the edges of the objects. This training method requires the following processes:

1. Offline training of image segmentation algorithms, that is, distinguishing pixels in the image by objects. This type of work is generally handled by specialized data labelers, who label different details in massive images according to needs;

2. Process image segmentation online and find suitable grasping points on manually marked objects.

This is a widely used method and the main driving force for the advancement of robotic arm grasping. The development of robotic arm technology is slow, but the image segmentation is progressing rapidly, which also pries the development of robots and other industries from the side.

Model-free

This training method does not involve the concept of "object". The machine directly trains the grasping strategy from the anpodal point, that is, the point that is possible to "grasp". This training method often allows the robot hand to try a large number of different types of objects and conduct self-supervised learning. Farm is one of the representatives.

Google Arm Farm

It is worth noting that for a robot arm, the difficulty of grasping objects of different shapes varies greatly. Even for objects of the same shape, the difficulty of grasping in different scenes varies greatly due to the influence of surface reflectivity and ambient lighting. There is still a long way to go from the laboratory to commercial implementation.

The development of high-precision cameras is the first step for robots to "perceive" objects.

In actual business scenarios, the most troublesome object is always the "next object". To truly integrate into the actual production system, only by having a smart brain and making flexible adjustments for different working conditions can the use scenarios of robots be broadened.

Reviewing Editor: Huang Fei