I've seen a robot screwing in a factory, but have you ever seen a robot stir-frying at home?
I believe there are many poor friends who have already swiped a video of a robot cooking a few days ago.
Tear open the lettuce package, cut the vegetables, beat the eggs, a set of operations are smooth, and it will even use the lid to block it when pouring greens into the pot.
Don't hide it from everyone, Shichao's first reaction when he saw the video was: Ah???
My impression of robot cooking is still the knife and noodle robot in the cafeteria.
If you hadn't seen it with your own eyes, no one would have dared to believe that it was a robot.
This robot is Stanford University's Mobile ALOHA robotics project.
ACCORDING TO THE VIDEO POSTED BY THE PROJECT LEADER, IN ADDITION TO BEING ABLE TO COOK, THE ALOHA ROBOT IS ALSO PROFICIENT IN VARIOUS HOUSEHOLD CHORES.
Watering the flowers, teasing the cat, making coffee, mopping the floor, washing the dishes, folding the quilt, and washing clothes, these chores are completely out of the question.
Especially the last shake of the pillowcase, I seem to see myself...
Anyway, after the video was released, it caused a lot of heat at home and abroad.
Netizens commented, " As long as this thing doesn't kill me while I'm asleep, I really need it "" Next time if the family urges me to get married and have children, I will send the video to the family group. ”
However, there are also doubts that the video has been accelerated and edited, as was the case with the Gemini demo video released by Google last time, and it remains to be proven whether it is true or not.
SOON, A VIDEO OF THE ALOHA ROBOT OVERTURNING WAS CIRCULATING ONLINE.
Breaking wine glasses and dinner plates without saying a word, crashing into cabinets with unrecognizable obstacles, burning pots...
The level of intelligence is compared to the flexible robot in the video a few days ago, which is like two "machines".
Interestingly, the person who broke the news of this video is an insider of the project team.
Over the weekend, Tony Z. Zhao, co-lead of Mobile ALOHA's project, posted a snapshot of the robot's failures on the X platform and clarified the misconception that the robot is "fully autonomous".
In fact, the ALOHA robot is not fully automated, but rather a hybrid mode of automatic + remote control.
Shichao carefully flipped through the Mobile ALOHA project on GitHub, and in autonomous mode, humans need to control the robotic arm to make a sample before the ALOHA robot can learn similar actions.
The various complicated housework in the previous video are all remotely controlled by someone in the back.
It's like this.
Moreover, it is not as long as a human demonstrates that a robot can learn 100% of the time.
It is also mentioned in the paper that the ALOHA robot has a 95% success rate in automatically wiping red wine stains, 80% in pushing chairs, and only 40% in fried shrimp.
The world's super blind guess, it may be that the video effect at the beginning was too amazing, one pass ten to hundred, which made the ALOHA robot bear the title of "nanny robot has landed".
But as Tony Z.Zhao tweeted, "Robots aren't ready to take over the world!" Today's robots are still far from the all-round nanny we imagined.
Therefore, instead of bragging about how good the ALOHA robot is this time, it is better to say that it has set a model for future robot research and development.
According to Nvidia scientist Jim Fan, data has always been the Achilles' heel of robotics.
The project team's definition of an ALOHA robot is "a low-cost, full-body remote operating system for data collection."
What does that mean?
Algorithmically, they developed an action learning algorithm, ACT, based on Transformer.
This algorithm allows the ALOHA robot to mimic a human's actions after a 15-minute demonstration of the action.
You can understand that the process of human control of the robotic arm is also the process of data collection. Then, the collected data is used to conduct a wave of "parody show".
For example, if you demonstrate the action of cleaning the glass to the robot 50 times in advance, there is a high probability that the training will be successful.
In addition to the clever design of the algorithm, the ALOHA robot also has the advantage of low cost.
In terms of hardware configuration, the project team installed two sets of robotic arms on the robot, as well as an AGV mobile chassis used to transport goods in the warehouse, which is equivalent to the hands and feet of the robot.
Judging from the cost list published by the team, the expenses are mainly focused on the mobile chassis and robotic arm, and the cost of other parts is not high.
Moreover, the cost of nearly $32,000 (equivalent to 227,000 yuan) is a drop in the bucket compared to the cost of nearly $2 million for the Boston Dynamics Atlas.
When the technology matures in the future, it seems that it is not impossible to spend more than 200,000 yuan to buy a robot for the elderly.
What's more, this Mobile ALOHA is directly open source, including hardware code, datasets, and tutorials.
Developers, you contribute a little, I contribute a little, maybe one day the all-round nanny robot will really land.
Including a few days ago, when the ALOHA robot was gaining momentum, DeepMind also released a series of updates on the robot while the iron was hot.
There is a robot company called Figure abroad, and their humanoid robots now even have the ability to learn independently, and they can learn to make coffee by watching a video.
Anyway, Shichao is quite looking forward to the development of robots this year.
In the world of artificial intelligence, there has always been a Moravec paradox.
In layman's terms, AI can do a good job in things that require logical reasoning skills, such as math problems, but its perception and motor skills are not as good as even three-year-old children.
But the emergence of ALOHA robots this time does have a little bit of a real sense of "intelligent" robots.
We can also look forward to what kind of sparks the combination of large models + robots can collide this year.