Hengyu West Wind comes from the Au Fei Temple
Quantum Position | 公众号 QbitAI
The robot company that broke Boston Dynamics' somersault patent for the first time in the world with an "electric robot" is no longer hidden!
More skills and techniques of its humanoid robots were put on the table.
In fact, it also does some delicate work: making coffee.
It is also the world's first humanoid robot to realize autonomous coffee latte art.
I saw him shaking his wrist from side to side holding the latte flower cup (milk bowl), unhurried:
Then you will get:
Then the matter was brushed away, and the merit and fame were hidden:
In addition to high-precision self-made latte art in cafes, it can also distribute late-night snacks to factory employees and act as a guide in indoor and outdoor scenes.
It can even help us walk the dog.
Officially, the company behind it, MagicLab, focuses on general robotics and AI technology, and I heard that it is still taking a global route.
Humanoid robots are just one part of their product line, with bionic quadrupeds, industrial quadrupeds, general-purpose humanoid robots, and more all part of the company's plans (and in the pipeline).
The landing scenarios targeted are also wide, covering household, industry, business, etc.
It is understood that the landing application of 3C scenarios is under development, and it is expected that trial operation will be carried out on the production line by the end of the year.
Performance testing of cleaning equipment on the production floor increased efficiency by 30% while reducing the need for manual intervention by 70%.
Self-developed servo motors and large models are also used
The MagicLab team focuses on the core technologies of general robots such as robot core torque motor joints, robotic arms, and foot-type robots.
The team believes that the ultimate form of a general-purpose robot must be a humanoid robot.
Technically, the team focused on solving both hardware and algorithmic challenges.
In terms of hardware, the team chose self-developed servo motors and core components to extend their capabilities.
The design of torque motor joint modules is dedicated to improving the power density ratio and torque density of motors, including harmonic module joints, linear module joints, and small and micro servo units.
At present, the team has a series of motors, covering a torque range from 10N.m to 500N.m.
Previously, the electrically driven humanoid robot completed the somersault thanks to MagicLab's new joint design.
The self-developed D190 joint module has a rated output torque of 150Nm and a maximum output of more than 525Nm, and can work at a speed of 110RPM.
The mechanical structure design of the robot also adjusts the weight distribution of the robot, concentrating most of the weight at the base of the thigh, reducing the inertia in the process of swinging the leg.
During jumps and flips, the robot can quickly adjust its center of gravity, maintain stability and make a successful landing, while maintaining agility and precise control.
Another advantage of self-development is that it can consider the collaborative optimization of software and hardware as a whole, use relatively low-cost hardware to give full play to the available operation and control capabilities, and continuously iterate to reduce costs.
In terms of algorithms, the team applied a multi-modal large model, based on the general knowledge ability of the large model, which can cope with the multi-task physical representation perception and objective physical fact understanding in complex scenes, and form human-like behavior decisions.
In terms of the whole body control model, the operation control model based on reinforcement learning and the operation model based on the skill library can fully decouple the environment and objects, its own executor, and the physical characteristics of the operation, so as to make human-like flexible, robust, and universal whole body control.
Finally, through the graphical interface, combined with the above two large models, users can easily design and deploy robot applications to achieve multi-modal human-computer interaction, scene task perception and autonomous decision-making.
Judging from the specific examples, the difficulty to be overcome by the robot coffee latte art is not only to be able to simply latte art, but also to make the robot reproduce the craftsmanship of the coffee master.
It needs to interact with people in an open space, which means that there is a lot of uncertainty about its spatial position and trajectory, as well as tools that need to identify different materials and sizes.
As the core component of the humanoid industry chain, the dexterous hand is also the hardware that the team focuses on developing, and the team has adopted a generational iterative strategy:
Mature generation, pre-research generation, reserve generation.
Sub-millimeter positioning accuracy and a load capacity of several thousand grams can already be achieved.
The base version is designed with 11 degrees of freedom, of which there are 6 active degrees of freedom. The thumb is distributed with two active degrees of freedom and can be used to bend and swing sideways, allowing the dexterous hand to mimic various movements of the human hand, including fisting, opening, pointing, pinching, fiddling, etc.
In addition, force control and fingertip tactile perception have been introduced, and precise control is achieved through the feedback of the torque sensor of the main body. Objects of any shape, size and material, even those that are fragile, sliding or deformed, can be easily grasped.
The team then plans to introduce haptic perception and design with a higher degree of freedom.
"M0-M4" that defines humanoid robots
After some conversation, the future plans and goals of the MagicLab team are still very clear.
In the face of the current development stage and the future, MagicLab divides the landing level of humanoid robot scenarios into five stages: M0-M4.
Let's break down the criteria for the division of these five stages.
The first is M0.
M0 is a laboratory product and is not oriented to a clear scenario.
From the perspective of R&D, many people may pile up some software and hardware at any cost, without any clear cost control, even if someone buys from the perspective of scientific research, there is no way to really commercialize.
In M1, the humanoid robot at this stage is more like an intern, doing some local applications for specific commercial and industrial scenarios.
For example, ask the humanoid robot to pour a glass of water, dance a dance, and make a cup of coffee with latte art.
From the perspective of R&D, the M1 stage can already optimize products and costs based on specific scenarios and specific tasks. However, the generalization is far from enough, so it is not a truly practical product, and it is easy to fall into the problem of commercialization.
M2, which can be defined as a skilled worker.
Products at this level need to be able to achieve a closed-loop scenario in specific commercial and industrial scenarios, that is, they can really do something for people.
The purchaser can calculate an economic account of the hardware and labor costs of the humanoid robot to judge the commercialization feasibility.
It is also at this stage that humanoid robots have really come to the door of commercialization.
MagicLab defines the M3 humanoid robot as "like a babysitter".
At this time, humanoid robots can already enter the home and complete tasks such as companionship, housework, and elderly care.
"However, we fully realize that before we can achieve this goal, we need to fully improve the generalization capabilities of large models and hardware." Zhou Yang, head of R&D at MagicLab, said that to reach this stage, the entire industry will need to spend a lot of time exploring.
The highest level, M4, is what MagicLab calls Siri in the physical world, which is a true general-purpose robot.
It can be used for multiple purposes – depending on the hardware level of the humanoid robot.
However, M4-level humanoid robots must be versatile and fully integrated into society, and may be able to go out to work and earn money in their spare time (??? ）。
The MagicLab team told us that the long-term goal is to make products at the M3 to M4 level.
At present, its own humanoid robots are currently in the transition stage from M1 to M2 evolution, and at the same time focus on promoting the landing of M2 level.
Get the robot out of the lab
MagicLab was founded in mid-December last year.
So far, it has grown into a team of more than 100 people. The members come from all over the world, and more than 80% of them are R&D personnel.
The team focuses on the core technologies of general robots such as robot core torque motor joints, robotic arms, and foot-type robots, and has also made small achievements at the algorithm level, such as operation control algorithms, navigation algorithms, vision and AI algorithms, etc.
At the same time, it has the ability of independent research and development and manufacturing robots, multi-scenario landing capabilities and mass production delivery capabilities.
In the communication with Zhou Yang, the head of R&D of MagicLab, and Ivan, the chief strategy officer, Qubit learned that the team is dedicated to one thing -
Efforts should be made to make robots truly "replace people" and achieve closed-loop landing in specific business scenarios.
Including when we asked about the advantages of the team, Zhou Yang also said:
In addition to deep technical research, we also attach importance to the power of technology × business, and constantly explore the application scenarios of technology and do a good job in product engineering.
We believe that with the mature mass production of general-purpose ontologies such as dexterous hands and humanoid robots, general-purpose humanoid robots will also be able to obtain positive feedback on commercial value.
The whole team very much hopes to use robot technology to combine the industry and amplify the value of robots exponentially.
In addition, because MagicLab felt that the imagination of a robot equipment company was too limited, they used "Robot+" to position themselves.
Just like the Internet+ back then, it brings technology to thousands of industries and solves practical problems.
— END —
QubitAI · 头条号签约
Follow us and be the first to know about cutting-edge technology trends