The newly born robot dog can walk by himself after rolling for an hour. The latest achievement of Wu Enda's first disciple

bright and quick From the Aofei temple
qubits | official account QbitAI

Now? , Let the mechanical dog roll by itself An hour , It can learn to walk ！

The gait looks quite modular ：

Can also carry Live in a big stick of fierce hatred ：

Even if I fell in all directions , Turn over and stand up again ：

So it seems , Training mechanical dogs is no different from training ordinary dogs .

This is it. UC The latest results from Berkeley University , Let the robot train and learn directly in the actual environment , No longer dependent on Simulator .

Apply this method , The researchers trained in a short time 4 A robot .

For example, what I saw at the beginning 1 A robot dog who learned to walk when he was young ;

also 2 Mechanical arms , stay 8-10 After hours of actual combat capture , Performance is close to Human level ;

And a small robot with computer vision , Groping by oneself 2 Hours later, , Can scroll smoothly to the specified position .

The study was conducted by Pieter Abbeel And so on ,Pieter Abbeel He was the first doctoral student of Wu Enda , Not long ago, he just got 2021 ACM Calculation Award （ACM Prize in Computing）.

at present , All of the software infrastructure for this approach is open source .

A place called “ Dreamer ” The algorithm of

The of this method pipeline It can be roughly divided into 4 Step ：

First step , First, put the robot in the real environment , collecting data .

The second step , Transfer the data to Replay Buffer. This step is to use historical data for training 、“ To sum up your experience ”, Use the collected samples efficiently .

The third step ,World Model Will learn from existing experience , then “ Cerebral repair ” Out of strategy .

Step four , And then actors and critics （Actor Critic） Algorithm to improve the performance of the strategy gradient method .

And then go back and forth , Apply the refined method to the robot , In the end, there is a “ Explore and learn by yourself ” The feeling of .

The specific term , The core link here is World Model.

World Models yes 2018 Year by year DAVID HA A kind of Fast unsupervised learning , To obtain the NIPS 2018 Of Oral Presentation.

Its core idea is that human beings are based on existing experience , Form a mental world model , Our decisions and actions are based on this internal model .

For example, when people play baseball , The speed of response is much faster than the visual information conveyed to the brain , Then the reason why the ball can be returned correctly in this case , Because the brain has made instinctive predictions .

 

before , be based on World Model such “ Cerebral repair ” Learning methods of , Google put forward Dreamer This scalable reinforcement learning approach .

The method proposed this time is based on this , be called DayDreamer.

（ It seems that he can be called a visionary ？）

 

The specific term ,World Model It is an agent model .

It includes a visual perception component , The image can be compressed into a low dimensional representation vector as model input .

There is also a memory component , Can be based on historical information , Predict the future representation vector .

Last , It also includes a decision component , It can be based on visual perception components 、 The representation vector of the decision component , Decide what action to take .

Now? , Let's go back to this UC The method proposed by Berkeley scholars .

It's not hard to find out , among World Model Learning Part of the logic is a process of experience accumulation ,Behavior Learning Part is a process of action output .

The method of this paper , It mainly solves the problems in robot training Two aspects The problem of ：

Efficiency and accuracy .

Generally speaking , The conventional method of training robots is reinforcement learning , Adjust the operation of the robot through repeated experiments .

But this approach often requires Very large Test of , In order to achieve good results .

Not only is it inefficient , And the cost of training is not low .

later , Many people have proposed to train robots in simulators , It can increase efficiency and reduce cost .

But the author believes that , The simulator training method is accuracy The performance is still not good enough , Only the real environment can make the robot achieve the best effect .

From the results , In the process of training robot dogs , Only the flower 10 minute Time , Robot dogs can adapt to their own behavior .

and SAC Compare the methods , The effect has been significantly improved .

During the training of the manipulator , This new method also overcomes the challenges of visual location and sparse reward , The training results in a few hours are obviously better than other methods .

Research team

It is worth mentioning that , Members of the research team who brought new results this time , It is also very eye-catching .

among ,Pieter Abbeel He is wuenda's first disciple .

He is now UC Berkeley professor of electrical engineering and Computer Science , Director of Berkeley robotics learning laboratory , Berkeley AI Co director of the Institute , Once joined OpenAI.

Not long ago , He also got 2021 ACM Calculation Award （ACM Prize in Computing）, In recognition of his contribution to robot learning .

meanwhile , He is still AI Robot company Covariant Co-founder of .

another Ken Goldberg, It's also AI Top experts in the field .

He is now UC Berkeley professor of Engineering , His research direction is reinforcement learning 、 Human computer interaction, etc .

2005 year , He was voted IEEE academician .

meanwhile ,Goldberg And an artist , yes UC Berkeley art 、 Founder of the Symposium on science, technology and culture .

Besides ,Philipp Wu、Alejandro Escontrela、Danijar Hafner Three people work together .

among Philipp Wu It's just UC A senior in Berkeley .

One More Thing

While watching the robot dog training video , We found that the researchers used Unitree Mechanical dog ,

This brand comes from Yushu technology, a Chinese enterprise , The Mavericks who have been on the Spring Festival Gala before , Also from his home .

and , Recently, Yushu robot dog conducted a collective Go1 Test video exposure , It is also popular abroad .

Address of thesis ：
https://danijar.com/project/daydreamer/

Reference link ：
https://worldmodels.github.io/