https://mp.weixin.qq.com/s/akZcJFJInI0KrCPQu9Hlxw
By 超神经
对于叠叠乐（ Jenga ）这种积木游戏，想必很多人都深有体会，因为稍有不慎，辛辛苦苦堆起来的积木塔就毁于一旦。这样的事情，交给 AI 和机器人来做，会怎么样呢？
It seems that artificial intelligence teams always like to find breakthroughs through games. The robotic arm developed by the MIT team also started with games in the three-dimensional world.
In Jenga, generally, the blocks are first stacked with three blocks into one layer, and the blocks are staggered to form a tower, and then the blocks are drawn from the lower part and placed on the top of the tower to create a higher block tower.
The Jenga game is a test of patience, balance, strength and so on. For many people, (especially students who are prone to shaking hands) this game is too difficult. The robot developed by MIT easily overcomes this task through detection, algorithms, push-pull, alignment and other operations.
Where is it sacred?
Human beings always say that they "hand shake", so the research of robotic arms is to complete some refined or high-risk operations. One of the project team members, Alberto Rodriguez, an assistant professor in the Department of Mechanical Engineering at MIT, pointed out that the key to this robot is that it perfectly combines vision and touch.
This technology is not only on the Science Robotics journal, but
also on CCTV’s midday news today.
But from the appearance, this robot is similar to some common application machines, like an ordinary mechanical arm, but it is equipped with a soft tooth gripper, a force-sensing wristband and an external camera, which is quite Yu gave it hands, touch and eyes.
At work, the gripper is used to operate the building blocks and can also feedback the sense of touch; the sensor wristband is used to control the force of operating the building blocks; the camera is used to collect visual images.
In addition to having these shapes that allow the robot to move the building blocks flexibly, the most important thing is to have a "soul" that is different from the previous robots-researchers use new algorithms to make it better at this task.
According to MIT researchers, this robot does not use traditional AI learning methods, but creatively uses hierarchical model dynamics to build a clustering learning model.
The advantage of this is that it no longer depends on a large amount of data, but can make real-time analysis based on the feedback data, while contacting and detecting, while predicting the plan of moving the next block.
How does it play Jenga?
In fact, the robot can handle the seemingly complex Jenga game, and cluster learning is the key.
The traditional way to solve this game is to collect all the relationships that occur between the building blocks, robots, and building block towers to calculate the best way. But this will obviously bring a huge amount of data, and the calculation difficulty is greatly increased.
In this research, the robot was chosen to imitate the way humans play games. The first is to try to label and cluster the data. Then judge the feasibility of the new operation by comparing with the marked data.
First, let the robot face a building block tower, randomly select the building blocks and push them out with a relatively small force. For each operation of pushing out and drawing the building blocks, the computer will record the corresponding visual and force data, and together with the operation result Mark it out.
It took about 300 attempts in this study, and enough data was accumulated, and then the data was processed. Clustering is used here, and operations with similar data and results are grouped into a group to represent specific building block behavior.
Different groups represent different degrees of operability, which is also the standard for measuring each operation. For example, one set of data represents an attempt by the robot on a building block that is difficult to move, while another set of data represents an attempt on a building block that is easier to move.
For each different data set, a simple model is given accordingly. Combining these models, the robot is equivalent to learning real-time learning.
Finally, the actual exercise can be carried out. When the robotic arm pushes out the building blocks, it uses the camera and wristband to receive visual and tactile information, and then compares the received feedback with the previous data. If the data corresponds to a good result, Just perform this operation, if there is a risk of collapse, give up this operation.
Not just Jenga
MIT researchers pointed out that although robots have been able to play this game in their research, if they are used to compete with human masters, it is estimated that some improvements are needed. Because in this research, the AI robot focuses on solving the problem of physical interaction, which solves the problem of whether this building block can be extracted and placed on it. But the Jenga game still requires some strategies, which involves considering and analyzing the associated steps.
But the MIT research team obviously did not have this idea. Perhaps for them, creating a master of Jenga does not have much value. According to team researcher Rodríguez, the technology is being considered in actual work environments, such as in the manufacturing of robots on assembly lines.
Anyway, we often say that our hands are shaking. Then let professional "people" do these professional things. Let's wait to be overwhelmed by the New Year's food.
Super Nervous Encyclopedia
Clustering
Clustering is a machine learning technique for grouping data. Generally, data objects are grouped into multiple classes or clusters (Clusters), so that objects in the same cluster have a higher degree of similarity, while objects in different clusters are quite different.
For a given set of data points, a clustering algorithm can be used to classify each data point into a specific group. In theory, data points belonging to the same group should have similar attributes and/or characteristics. Clustering is an unsupervised learning method.
