Please select To the mobile version | Continue to access the desktop computer version

Nujin

 Forget password?
 Register

WeChat login, one step

View: 4291|Reply: 52

Artificial intelligence algorithm teaches robots to learn to walk: starting from scratch, it takes two hours

[Copy this link] Choose your language

3760

Threads

710k+

Posts

3639

Credits

白领

Rank: 4

Credits
3639
Post time 2019-1-26 06:25:25 | Show all posts |Reading mode

"Tianjin Robot"is the official subscription number ofTianjin Robot Industry Association, follow us, Learn about the robot industry information inside and outside Tianjin and at home and abroad.Units within the association can send the latest developments, achievements and activities toweixin@TJRIA.com.

In December 2018, scientists from the University of California at Berkeley and Google Brain developed an artificial intelligence[jzsjiale_guanjianci]artificial intelligence,=2[/jzsjiale_guanjianci]system that allows robots to learn to walk. This result was published on the preprinted website arXiv.org, and the paper is entitled "Learning to Walk via Deep Reinforcement Learning".

pVzGr1iBxl3Lx3bx.jpg

In the video material released by the researchers, the quadruped robot Minitaur tried to walk over a flat, gentle slope. At the beginning of the video, the time is displayed as 0, which is the beginning of the quadruped robot learning to walk. At this time, the quadruped robot Minitaur is like a toddler, shaking from time to time and stepping in place. It tries to open its "legs" to move forward, but its body stays in place "honestly", and the whole walking process progresses slowly. The change occurred in the 18th minute when Minitaur was learning to walk. At this time, it was able to move forward continuously, but its balance was slightly insufficient. In subsequent exercises, Minitaur's pace gradually became stable and rapid. In less than 2 hours, 54 minutes, 72 minutes, and 108 minutes, Minitaur basically learned to walk the gentle slope quickly and smoothly.
This is the whole process of the quadruped robot Minitaur learning to walk. The artificial intelligence algorithm[jzsjiale_guanjianci]artificial intelligence,=2[/jzsjiale_guanjianci]developed by scientists at the University of California, Berkeley and Google Brain "taught" this quadruped robot through familiar or unfamiliar terrain.

During the entire training process, researchers need to "manually" "please" the robot that has reached the end of the gentle slope back to the starting point of the gentle slope to restart a new round of practice. This manual resetting process is a bit cumbersome. However, judging from the results, this 2-hour learning process is really efficient, and many netizens commented that "AI is really a good teacher."
The concept of "reinforcement learning" is often mentioned in the field of artificial intelligence, which is an artificial intelligence method that uses rewards or punishments to achieve specific goals, with the purpose of obtaining a strategy to guide actions. For example, in the game of Go, this strategy can guide where to place each move according to the board situation. While the quadruped robot Minitaur is learning to walk, this strategy can tell the robot how to walk next based on terrain and other factors.
Reinforcement learning will start with an initial strategy. Usually, the initial strategy is not necessarily ideal, as the quadruped robot Minitaur showed when it first learned to walk. but. In the process of learning, the quadruped robot Minitaur, as the main body of decision-making, will interact with the environment through actions and continuously obtain feedback, that is, reward or punishment, and adjust and optimize the strategy based on the feedback.

Reinforcement learning is a very powerful way of learning. Continuous reinforcement learning can even obtain better decision-making mechanisms than humans. The best example is Alpha Dog. In 2016, Google's AlphaGo program trained through deep learning defeated the former Go world champion Lee Sedol by a score of 4-1. Its improved version defeated the world's number one Chinese chess player Ke Jie in 2017. Its shocking game ability was trained through intensive learning.
But reinforcement learning also has its limitations. It requires a lot of data and in some cases tens of thousands of samples to get good results. This requires the quadruped robot Minitaur to perform multiple trainings like the Alpha Dog, but too much training may cause damage to the quadruped robot.
Therefore, this "learning to walk" study chose the "upgraded version" of reinforcement learning-the deep reinforcement learning method, which combines the perception ability of deep learning with the decision-making ability of reinforcement learning. This method can be directly controlled based on the input image and is an artificial intelligence method that is closer to the way of human thinking.

In the words of the researchers, in order to "make it possible for a system to learn motor skills without simulation training", they adopted a reinforcement learning framework called "maximum entropy RL". The maximum entropy RL can optimize the learning strategy to maximize the expected return. In this framework, the artificial intelligence agent continuously finds the best course of action by extracting certain actions from the strategy and receiving rewards.
The researchers said, “As far as we know, this experiment is the first case of deep reinforcement learning algorithm that directly learns under-driven quadrupedal movements in the real world without imitation and pre-training.
In May 2018, researchers from the same research group published another research paper on the quadruped robot Minitaur on arXiv.org. At that time, researchers used deep reinforcement learning methods to make Minitaur learn quadruped sports from scratch, and finally realized trotting and galloping.

MLjzozz9MqLtS5lc.jpg
Part of the content comes from the Internet, and the copyright belongs to the original author. If it constitutes an infringement, please contact us and we will deal with it as soon as possible.
Reply

Use magic Report

3627

Threads

710k+

Posts

3661

Credits

白领

Rank: 4

Credits
3661
Post time 2019-1-26 09:48:54 | Show all posts
那一夜我抱着你,在你耳边叫你戴上那玩意,你说不戴的感觉才够爽,现在是安全期,没事……可不戴头盔交警抓着咋办
Reply Agree Disagree

Use magic Report

3788

Threads

710k+

Posts

3673

Credits

白领

Rank: 4

Credits
3673
Post time 2019-1-26 09:48:12 | Show all posts
挨骂也是幸福~~~
Reply Agree Disagree

Use magic Report

3498

Threads

710k+

Posts

3636

Credits

白领

Rank: 4

Credits
3636
Post time 2019-1-26 09:48:11 | Show all posts
相比他连说拜拜的想法都没了哈哈
Reply Agree Disagree

Use magic Report

3750

Threads

710k+

Posts

3659

Credits

白领

Rank: 4

Credits
3659
Post time 2019-1-26 09:44:56 | Show all posts
如果寒暄只是打个招呼就了事的话,那与猴子的呼叫声有什么不同呢?
Reply Agree Disagree

Use magic Report

3549

Threads

710k+

Posts

3656

Credits

白领

Rank: 4

Credits
3656
Post time 2019-1-26 09:43:48 | Show all posts
如果寒暄只是打个招呼就了事的话,那与猴子的呼叫声有什么不同呢?
Reply Agree Disagree

Use magic Report

3894

Threads

710k+

Posts

3663

Credits

白领

Rank: 4

Credits
3663
Post time 2019-1-26 09:43:35 | Show all posts
必须顶
Reply Agree Disagree

Use magic Report

3560

Threads

710k+

Posts

3640

Credits

白领

Rank: 4

Credits
3640
Post time 2019-1-26 09:43:34 | Show all posts
原来还有这么多内幕啊,长见识了,呵呵
Reply Agree Disagree

Use magic Report

3445

Threads

710k+

Posts

3635

Credits

白领

Rank: 4

Credits
3635
Post time 2019-1-26 09:42:57 | Show all posts
辣妹打电话叫出租车。对方:小姐,请问你待会儿穿什么衣服?辣妹:红色超短裙!对方:那到哪里呀?辣妹:到大腿啦!
Reply Agree Disagree

Use magic Report

3806

Threads

710k+

Posts

3652

Credits

白领

Rank: 4

Credits
3652
Post time 2019-1-26 09:42:39 | Show all posts
人生感悟:爱她,但别怕她,你们是恋人,也是朋友,她要的不是宠物,这样的感情,走不长远。
Reply Agree Disagree

Use magic Report

3561

Threads

710k+

Posts

3663

Credits

白领

Rank: 4

Credits
3663
Post time 2019-1-26 09:41:51 | Show all posts
日照香炉生紫烟,疑是熊猫在烧香
Reply Agree Disagree

Use magic Report

 
Online Customer Service
点击这里给我发消息
Technical Advisory Hotline
Wechat:gwailee
QQ:395508196
EM:hi@niuresou.com

Scan WeChat, follow the public account to lead Nujinniu Stock Niu Ce

QQ|Baidu Statistics|Archives|Mobile|Black House|Search management|Game management|Nujin ( ICP证:沪B2-20210491 | ICP备案号:沪ICP备20013403号 ) Choose your language

GMT+8, 2021-6-23 10:31 , Processed in 2.435546 second(s), 3970 queries .

Powered by Nujin X3.4

© 2001-2017 Nujin Inc.

Quick reply To Top Back to list