After the Stanford shrimp-frying robot exploded last week, another coffee-making robot became popular. What’s great about it is that by watching human demonstration videos,Only 10 hours of end-to-end training, you can learn and complete this task completely independently. Of course, more precisely, operating a coffee machine:
I saw it open the cover, put in the coffee pod, and press the start button. The whole process was completed in one go.No need for any remote control.
Soon, a cup of coffee is ready to go and enjoy:
Note that the above demonstration video does not have any acceleration processing. This is the actual operating speed that the robot can currently achieve.
In addition to these, it also hasAutonomous error correctionFunction:
If the coffee pod is not placed correctly, it can adjust itself without any human reminder.
In this regard, the former Google DeepMind researcher (also the developer of the robot) directly forwarded and liked it, and emphasized again:
All actions, including error correction, are completely autonomous.
Company founder Brett Adcock said this is the ChatGPT moment for robotics.
Regardless of whether everyone agrees with this statement, there is no need to say that netizens are impressed by its training speed.
10 hours is really excellent. Fast forward to using the grinder and the French press. It is estimated that the coffee machine in the video will soon be out of work. (Manual dog head)
So, what is the specific origin of this robot?
10 hours of end-to-end training to make coffee
The robot above comes from a commercial company called Figure.
Figure is headquartered in the United States and founded in 2022, specializing in general humanoid robots.
The founder, Brett Adcock, graduated from the University of Florida in the United States. At the age of 26, he founded an online talent market website, which was later acquired by his peers for US$110 million. He then founded an aerospace company that produces all-electric vertical take-off and landing aircraft and went public for US$2.7 billion.
Today, the robot company Figure is also attracting huge amounts of money. It first received US$70 million in Series A financing in May last year, and two months later received a US$9 million investment from Intel.
In October last year - about a year after its establishment, Figure released its first humanoid robot, code-named Figure01.
This is how it walks:
The coffee making demonstration we saw today was completed by it.
It only took 10 hours of training to learn this skill——
Figure01 uses an end-to-end neural network. This side receives videos of humans making coffee, and the other side can output action trajectories, allowing the robot to imitate and ultimately complete autonomous operations.
By analogy, to let it learn other tasks, it only needs to input the corresponding video.
As for the specific implementation details, the official did not disclose it.
However, aside from completing the training in 10 hours, it is no longer difficult for the robot to learn to make coffee.
Its core is imitation learning, and VIOLA, which was selected for CoRL’22 (Robot, Learning Conference), can do it (it is an object-centered imitation learning framework, based on Transformer for reasoning, good at long-range tasks, and has a performance 45.8% higher than the most advanced imitation learning algorithm);
HYDRA, which comes from Google and was released in June this year, also specializes in this. It is good at various coarse-grained and fine-grained controls and can switch freely.
AME from Stanford University is based on Waypoint and is unambiguous for tasks such as making coffee, but it is much slower.
There are also NVIDIA's HITL-TAMP, MimicGen and other research results, which are related to robot imitation learning, so I will not introduce them one by one.
Who can successfully challenge Steve's coffee test before 2040?
Although the performance of Figure01 is very good (for example, it is very fast), many netizens still commented:
This is still a little far from the imagined coffee-making robot.
For example, can it pick up a cup, put it under the coffee machine, add cream and sugar after brewing, put the cup on a tray and hold it in front of a person?
In fact, Apple co-founder Steve Wozniak once proposed a coffee test to evaluate the autonomous capabilities of robots.
It requires the robot to start by entering a strange home, then find the kitchen, identify tools (such as coffee machine, kettle) and materials (coffee beans, sugar, milk, etc.), and finally make a cup of coffee, and the whole process does not take more than 20 minutes.
This test tests the robot's ability to navigate in unknown environments, recognize objects, operate tools and materials, and follow human commands.
Someone has initiated a vote and asked whether such a robot can be born before 2040.
As a result, 89% of people voted yes.
Do you think there is hope?