Jin Lei Fengse from Ao Fei Si
qubit | Public account qbitai
Stanford’s all-around housework robot mobile aloha, big! turn! ! car! ! !
You thought it would be easy to drink red wine, but in fact it is like this:
I will spread it all for you, and break a cup...
You thought it could transform into a chef and cook skillfully, but in the end, it will give you a stir-fry Bottom of the pot :
mobile aloha's rollover collection is more than that.
For example, I just finished frying the shrimps in the pot. Oops, I accidentally missed it:
Even if the little brother rushed forward, he didn't stop the "tragedy" from happening (it seemed to have burned his hand).
This scene really looks like Mrs. Zhuang throwing a bowl...
Mobile Aloha, which was still on the "shrine" yesterday, was exposed to so many "clumsy" behaviors overnight, which also attracted the attention of many netizens. .
However, this time, even if faced with irrefutable evidence, netizens’ painting style was uncharacteristically:
It’s not perfect, but it’s cute.
There is always room for error.
The most important thing is:
is relieved. (Manual dog head)
What on earth is going on?
Stanford team exposes "scandal"
It turns out that this video of the robot overturning was posted by Tony Z. Zhao, the author of Stanford Mobile Aloha.
And he also bluntly said:
Robots are not ready to take over the world.
This rollover video is exactly what the robot did in autonomous mode . In the author's words,
is the "stupidest mistake" of and .
After all, in addition to the few examples we just showed, mobile aloha can't even put the pot in the cupboard:
The fried shrimps stick to the pot and can't be poured out, and even the bowl can't be found:
Get one The pen can't find the right position to start:
Faced with the collection of failures, the author joked:
This is my favorite video so far, (but) when the robot makes a mistake in front of you, you won't feel So interesting.
Indeed, after all, my hands were burned...
However, the author revealed that there should be another reason for this video today.
The video of mobile aloha’s god-level scene in the past two days has indeed attracted a lot of attention, but many people mistakenly thought that it was done in autonomous mode.
But in fact, mobile aloha uses a hybrid model and is not completely autonomous. The author also calls on netizens to read the paper and code carefully while eating.
It is worth mentioning that the author also quoted the 2015 Boston Dynamics atlas humanoid robot "Rollover Collection" and paid tribute to it.
Perhaps this is what Nvidia scientist Jim Fan said:
One step at a time.
After learning 50 times, the success rate can reach 90%
In the past two days, the mobile aloha team released three popular videos in a row, showing the robot's agile and dexterous housework capabilities, which stunned netizens.
Including making a Manchu-Han banquet (breaking eggs, flipping chicken, and other details are all at your fingertips):
Setting pillowcases and laying sheets:
Watering flowers, mopping the floor, opening bottle caps, and even teasing cats:
That's like a human being, There is a living room and a kitchen below.
However, most of them are real people controlling , such as the above.
For a more intuitive view, you can look at the following animation of pulling out paper and cleaning the glass. There is a human 1:1 demonstration directly behind it:
However, for some relatively simple tasks, such as this single fried shrimp:
and cleaning the pot. It can learn to do things like returning the dining chair, calling and taking the elevator, wiping the table, etc., with only a small amount of teaching from a real person, and then can operate independently without humans.
Specifically, the author said that the above simple actions only need to be learned 50 times to achieve a 90% success rate -
After testing, mobile aloha can wipe up the spilled wine 9 times in a row and call 5 times in a row. There will be no mistakes in every elevator and a certain degree of stability can be maintained.
In addition, it is also anti-interference. When finishing placing the pots in the cabinet, the experimenter kept throwing debris in front of it, but it did not affect its performance at all:
It was not visible at all during the training. Chair? It can also accurately identify and complete the homing task.
So, how did the author enable mobile aloha to achieve autonomous tasks with only 50 demonstrations? The most important thing about
is to imitate learning through act or diffusion strategy, and then jointly train and on the robot system together with static operation data.
With this joint training method, the performance of the robot can be significantly improved, especially for tasks that require precise operation.
Finally, let me once again introduce the results of this robot from Stanford University:
was officially released at the end of March this year and has gone through 8 months of iteration and 2 months of testing.
has a total of three authors, two of whom are Chinese doctoral students majoring in computer science at Stanford (the last one is a supervisor):
At that time, the robot was already able to use tools to complete various precise and detailed tasks, but only Can be in a fixed position:
Of course, the back is also controlled by a real person.
As its name aloha stands for "ow-cost pen-source rdware system", this robot focuses on open source and low cost: all software and hardware designs of
including code and data are released together, and building this system " It only costs 32,000 US dollars (approximately 227,000 RMB). The author also made a list of the specific hardware required. Interested friends can DIY according to this.
The first year of the robot?
Almost at the same time as Stanford’s explosive robot, Google also released its latest research results, and published three of them at once:
One is to increase the robot’s decision-making speed by 14%, while maintaining the same operation accuracy and increasing it by 10.6 % improved model;
One is a new framework that specializes in generalization capabilities. It uses a new method to increase the success rate of robots in completing never-before-seen tasks from 29% to 63%;
and one that can accommodate 20 robots at the same time. The receipt collection system will be used to speed up the training of robots’ ability to understand human instructions.
These new results are all used to upgrade Google's large robot model rt-2.
Compared with Stanford's mobile aloha, Google's rt-2 is still colder, but all its effects are completely independent.
In addition to these two companies, Li Feifei’s team has also been following up. Its robot system called voxposer can also understand human speech and complete various instructions without additional training.
This makes people think of many people’s prediction that “2024 will be the year of robots”:
Do you think it will come true?
reference link:
https://twitter.com/tonyzzhao/s
