Shulou(Shulou.com)06/02 Report--

If we want to exercise a human child's most basic motor intelligence, such as grasping and lifting, or housework such as folding quilts and clothes, there is little need for guidance.

Most of the time, as long as the baby is placed in the situation of daily life and allows him to interact with the environment and objects around him, most of them will be able to automatically master the ability to judge space, gravity, coordination and so on.

In contrast, robots are more "spoiled".

At present, most robots are unable to adapt themselves and learn general motion ability from complex environments.

This brings a problem: the robot can only "clumsily" complete some of the actions first programmed by the programmer, and use a single object to complete a single skill, this particular task and prop is its world. For example, if you can hand over a cup, you can't make a quilt.

This means that we may have to design thousands of robots just to deal with a specific task. This is so stupid.

However, it is a big project for machines to learn to perceive the world on their own and act accordingly according to changes in the environment.

Recently, the University of Berkeley has developed a new algorithm, reinforcement learning based on visual model, which makes omnipotent robots possible.

In other words, meta-motor intelligence, which originally belonged solely to human beings, is also expected to be turned on in machines.

When robots have the ability to master general skills and internalize them into "experience", and can flexibly perform a variety of similar tasks without having to relearn or program every time, the prospect is obviously a pleasant surprise.

So, how on earth did such an amazing thing come true?

How does the new algorithm guide the machine to work?

In a nutshell, this algorithm can make machines learn a large number of diverse images independently through a prediction of Mexico and the use of untagged sensory data sets, thus making it more flexible to predict and judge when completing tasks. In this way, it can perform many different tasks on a variety of different objects without having to relearn for each object or task.

In the eyes of researchers at the University of Berkeley, being able to acquire this universal motor ability in a single mode is a basic manifestation of intelligence.

So, how on earth does this method guide the robot to complete the work?

First of all, the researchers create a large and colorful data set for the robot, which is not limited to a particular object or skill.

Then, the robot is equipped with various sensors that can sense image pixels (vision), arm position (self-perception) and send motor instructions (action).

After these preparations are completed, the two robots are allowed to collect data and learn independently in the resource pool at the same time, and share them in real time.

Because the two robots can share each other's feelings and data, they have the ability to predict the next path of arm movement, thus making the range of motion more flexible and flexible to complete a variety of tasks of operating different types of objects.

For example, they let a machine move rigid objects such as Apple and flexible objects such as folding clothes, and the robot performed well.

Moreover, even in the face of goals that have never been seen before, although the predictions made are not as perfect as human beings, they can still effectively accomplish the assigned task.

For example, in the picture below, the task given by the researchers is to put the apple on a plate, with the plan made by the robot in the middle, and then the execution.

This algorithm model is absolutely unique in terms of ideas. For a long time, planning the course of action based on the expected results and iterating and improving the plan in real time according to different steps and observations has always been a unique skill for human beings to deal with accidents in a complex world. Today, robots are also expected to learn such "high-intelligence" games, and the changes brought about in application are really exciting.

Rebuilding the coordinate system of machine efficiency: the application scenario of the new model

The real environment is complex and diverse, which improves the adaptability of the robot to the environment and makes the task execution more flexible. this new algorithm does bring a lot of imagination to the robot application.

The biggest highlight is that the assumption of the function of the machine is more in line with the reality. General capabilities can be easily migrated to different tasks, greatly reducing the number of algorithms that need to be developed and deployed to accomplish specific tasks.

For now, it seems that the new model will revolutionize robot performance in at least the following areas:

1. Customer service robot. Most customer service robots are not adaptable to the environment, so programmers need to take all kinds of situations into account, and some even operate manually in the background. However, with the general model algorithm, the robot can learn independently in the interaction with human users, learn to solve some open problems, and become more autonomous and flexible.

two。 Medical machinery. At present, medical robots can only be used as an extension of doctors' hands to assist in surgical tasks. In order to perform high-precision surgery independently and help reduce the labor of doctors, medical robots must be able to perceive the spatial location of the surgical site and handle more elaborate and complex operations. The new model obviously provides more possibilities.

3. Industrial machinery. Industrial robots already have a certain degree of versatility and adaptability, but they often need to reprogram according to the changing needs of their working environment, or change different operators to perform different tasks, which will bring certain costs. If the new algorithm is really applied, the cost and efficiency of industrial production will become lower.

4. Personalized video generation. In addition to motion perception and understanding prediction in the real world, the algorithm also has a lot of room in the field of video generation. For example, the system can learn independently through a large number of unmarked video resources, identify and imitate the characters in the video, and make it possible for AI to tailor highly anthropomorphic videos.

There are many machine applications that have mastered general skills, and there are broad commercial prospects behind them. after all, efficiency is the original intention of human beings to invent machines.

What other challenges need to be addressed before tasting the fruit?

Having said so much, I feel that the implementation of the new algorithm is not difficult, and the application side also has enough support. Will it become a reality soon?

At present, before "picking peaches", the algorithm still has some special restrictions, which may hinder its practical application.

First, a large amount of training data is needed. The prediction of the actual executable operation of the machine depends entirely on the large and diversified data set.

In order to enable the machine to imagine and simulate the next pixel motion based on predicting the motion distribution of previous frames, the researchers introduced 59000 robot interactive data sets for large-scale self-supervised learning.

How to obtain huge and high-quality data resources under cost control may become the number one threshold for the landing of the algorithm.

The second is a series of problems brought about by unsupervised learning. For example, because there is no label and reward mechanism for the training data, how to ensure that the robot can understand and accept the specified task, and then act on the basis of results, there are a lot of unknowns in reality.

Berkeley's solution is to set up a self-monitoring algorithm that keeps the machine interested in the target, keeps tracking and retries until it succeeds. However, whether the output can be stable or not, more supplementary research is needed.

For example, the predictions made by unsupervised machines cannot be explained by independent experience, it may not be the optimal transmission scheme, and it may also bring unpredictable risks because of the "black box".

When we do not understand how our "meta-intelligence" works, how can we ensure that we can control the machine's "meta-intelligence" well?

All in all, although this algorithm is surprising, it is not perfect. The imagination is beautiful and practical, but there is still a long way to go from the lab to the commercial scene.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.