Author: Sophia
Internet of Things Think TankCompilation and Release
At 10:10 this morning, "Wild Iron Man", "Huawei Genius Boy", "Resigning to start a business with an annual salary of one million" - Mr. Zhihui, wearing these legendary auras, brought his first product show since his establishment.
Zhiyuan’s embodied intelligent robot “Expedition A1” is released!
Judging from the barrage before the opening, most people are looking forward to it~
After Mr. Zhihui came on stage (still a little nervous), he gave everyone a full and informative speech in 30 minutes.
According to reports, the intelligent robot "Expedition A1" integrates various advanced ontology control, perception, cognition and decision-making intelligent technologies, and is based on a series of innovations such as the very cutting-edge large language model in the current AI field, as well as its self-developed visual control model. technology will have excellent intelligent interaction and task execution capabilities, and can autonomously complete various movements and operations in various complex scenarios in the future.
This press conference was also the first time the brand of Zhihui Jun’s startup company, Agibot, was exposed to the public.
Why use this name? Because "Zhi" represents artificial intelligence, and "Yuan" is the pictogram of a person with two legs below, its English name Agibot means general artificial intelligence robot.
Therefore, the original intention & mission of the company Agibot is to "create robots that are as flexible and intelligent as humans, so that intelligent machines can create unlimited productivity", and ultimately promote the entire human society to move towards a more intelligent, efficient and sustainable world. A continuing future.
At the end of the press conference, the robot itself also appeared on the stage, which brought a perfect curtain to this debut.
How awesome is the "Expedition A1"?
There is a classic AI joke on the Internet, as shown in the picture below: 
We want the AI to cook, clean the room, wash and hang clothes, throw away garbage, shovel shit, etc., but what the AI is actually doing is chatting, painting, etc. Writing, composing, gaming and more…
If you want a robot to actually do the things on the left in the picture, you must achieve a deeper integration of AI and robotics technology, which involves a series of complex technical challenges such as environmental understanding, contextual reasoning, object recognition, motion planning, etc.—— This is not a hardware or software problem, it is a systematic collaboration problem of all aspects of software and hardware algorithms.
To this end, the Zhiyuan team has made staged progress in both software and hardware.
Humanoid robot body technology
Let’s first look at the hardware part, which is the humanoid robot body technology.
Why humanoid?
The answer is actually very simple, because the entire physical world in which humans live is designed for the human form, including various scenarios, tasks, equipment, and tools that are tailor-made for humans. Therefore, the humanoid body structure and abilities can obviously help robots Better integrate into the environment of human life.
Of course, Mr. Zhihui also admitted——
Humanoids are not the only solution for general-purpose robots. In some specific structured scenarios, robots on other platforms may have higher efficiency and usability, so they have also made various modular designs for the robot body.
But in the long run, starting from first principles, Zhihui Jun's team still believes that humanoid robots are the most common form in the human environment, and they are also the ultimate robot that can enter thousands of households in the future and hopefully enter everyone's lives. the ultimate form.
In order to realize the robot's smooth basic walking ability and up-and-down coordination ability, in terms of body hardware design, the Zhiyuan team equipped the robot with a total of 49 components including harmonic integrated joints, linear push rods, brushless planetary servo, coreless motor, etc. degrees of freedom, various types of actuators.
For a robot that moves very accurately and sensitively, its joints need to have a series of characteristics: small size, light weight, high enough power density, high enough energy utilization efficiency, high response bandwidth, impact resistance, etc...
Then, combined with the optimal design of the robot's overall structure design and control system, we can ensure its efficient execution.
Therefore, this core joint is not only the key to making humanoid robots more flexible and free, but it is also one of the important thresholds for future Zhiyuan to achieve large-scale mass production and low-cost manufacturing.
To this end, the Zhiyuan team independently developed and designed a special joint called PowerFlow, which weighs only 1.6kg. It integrates a liquid cooling circulation heat dissipation system into the small size and has a peak torque of up to 350Nm.
Another very important core component is the Zero Skill Hand.
The Zhiyuan team has independently developed a high-degree-of-freedom dexterous hand, which currently has 12 active degrees of freedom and 5 passive degrees of freedom. All drivers are built-in, and the cost can be as low as less than 10,000 yuan.
At the same time, because the A1 robot application will target some very precise operating scenarios, the fingertips of this dexterous hand are also integrated with vision-based fingertip sensors, which can distinguish the color, shape, and even material of the operating object, and can be based on various Algorithm data fusion is used to approximate the sensing effect of a tactile pressure sensor.
In addition to these two core components, the Expedition A1 is also equipped with a series of sensing components and computing systems, as well as supporting embodied intelligence frameworks. The efficient combination and organization of all these parts make it a highly intelligent and powerful robot development platform.
With a very good hardware platform as the foundation, the operation control algorithm is also a very core key. According to Zhihuijun, in just a few months of R&D practice, his team has rapidly iterated three versions of the operation control algorithm. It can be said that it already has the core algorithm research and development capabilities of the first echelon in the industry.
In order to make the expedition AI possible, the Zhiyuan team hopes to control the cost of the entire robot hardware within 200,000 yuan in the future. I wonder if Musk, who is holding the Optimus Prime robot, is panicking.
Embodied Intelligence Technology
If a robot wants to be mass-produced at low cost and produce practical application value, the hardware itself is only a prerequisite. What is more important is the robot's AI brain, which is the software part.
In recent years, with the rapid development of large language model technology, Zhihui Jun believes that it has become possible for robots to autonomously perceive the environment, understand task action choreography, and autonomously complete a set of processes.
In the view of the Zhiyuan team, the greatest value of large models in the field of robotics is reflected in two points:
The first point is "huge prior knowledge base & strong general understanding ability".
For example, when we need to ask the robot to throw the garbage on the table into the trash can next to it, we no longer need to tell it what the garbage is and what the trash can is. All these data are concepts that the pre-trained GPT model has learned from massive IoT data. Therefore, the emergence of large models can enable the robot to better generalize its original capabilities to scenes it has never seen. .
The second point is "complex semantic multi-level reasoning ability, the so-called thinking chain".
As shown in the picture below, if we need to instruct the robot to "put the blue square under the red square", then the robot has to understand what gravity is about, because things will not float in the air, how can he perform a multi-level operations to accomplish such a complex task?
Therefore, before the arrival of the era of large models, robots were all specialized equipment, and we needed to debug and deploy each task in a targeted manner. Now, using the various general knowledge abilities and reasoning abilities of large models, we can see solutions These problems finally lead to the dawn of universal robots.
Based on this, Zhiyuan has set up the "Embodied Intelligent Brain" EI-Brain framework, which divides the robot's thinking system into four levels, namely the cloud superbrain, cerebrum, cerebellum and brainstem.
After talking about technology, Zhihui Jun also used a video to show the unlimited potential of humanoid robots in various scenarios.
It is reported that the "Expedition A1" will first be used in new energy manufacturing and 3C manufacturing, and has already been docked with leading companies. However, the specific brand name has not yet been revealed.
The history of "genius boy" leaving his job and starting a business
Public information shows: Zhihuijun, whose real name is Peng Zhihui, graduated from the School of Information and Communications Technology, University of Electronic Science and Technology of China in 2018.
After graduating in 2018, he worked as an algorithm engineer in the AI laboratory of OPPO Research Institute. He was later invited to join Huawei through the Genius Youth Program and engaged in research related to Ascend AI chips and AI algorithms, with an annual salary of one million.
Being able to join the "Huawei Genius Youth Program" is already a good example of its strength. This program has extremely high requirements for applicants: those who have published a large number of papers in top international journals and have a high number of citations; those who have outstanding and abundant commercial results in research directions; those who have placed in well-known competitions will be given priority. In contrast, the recruitment process is also very strict. Generally, you need to go through about 7 rounds of processes, and finally pass the interview with Huawei's president.
Unlike other mysterious geniuses, Zhihui Jun is a big V who is active on the Internet. Many of his researches have also been published online. I believe that many technology enthusiasts have seen his videos on Station B. .
For example, this hard-core bicycle can not only stand stably on its own, but can also walk automatically, avoid obstacles, and even turn on its own...
There is also Iron Man's robotic arm. Zhihui Jun remotely controlled it to sew up a broken grape, but the grape was not damaged by the robotic arm.
After the "grape bandaging", the originally cracked small opening became like this:
And Zhihuijun "Geve" shared the self-made "wheel-footed" robot Nezha on station B in April this year. It is simply a combination of "national myth" and "cyber machinery" ~
Netizens marveled at the terrifying thing about Mr. Zhihui: One person completed all the details, modeling, algorithms, circuit design, driver design, APP development, and his knowledge reserve and hands-on practical ability were terrifying!
Zhihui Jun, who has accomplished many pioneering achievements, is relatively modest. He himself has expressed in many interviews, "I feel that I am neither a genius nor a teenager anymore." He is just a geek who loves to mess around.
In December last year, Zhihuijun officially confirmed his departure from Huawei, saying that he would start a new career next. Although Huawei has tried to retain him internally, it is obvious that it is difficult for a genius of this level to be a "screw" in a big factory.
In February this year, Zhihuijun’s startup company Agibot was officially established.
It is reported that Agibot is a company focused on the development of general humanoid robots and embodied intelligence. It is committed to the deep integration of A1 and robots and long-term investment in creating competitive robot products and ecosystems.
So far, Agibot has completed its third round of financing, and its market valuation has been revealed to be around one billion US dollars.
A few months ago, Mr. Zhihui also posted a recruitment post on Weibo, saying that artificial intelligence is ushering in a new round of industrial upheaval and is accelerating towards the Holy Grail of AGI. The combination of LLM + Robotics is an inevitable trend, and robots will In the near future, we will see an inflection point in the explosion of application scenarios. We hope that more people will join their young team and embrace the starry sea of the AI world.
Are embodied humanoid robots the future?
Judging from the new products released today, Zhihui Jun’s chosen direction is embodied humanoid robots.
The so-called embodied intelligence is Embodied Intelligence, or EI for short. Turing first proposed this concept in the paper "Computing Machinery and Intelligence" in 1950. Embodied intelligence allows machines to interact with and perceive the environment like humans, plan, make decisions, act autonomously, and have the ability to execute. It is also considered the ultimate form of AI.
Li Feifei, a well-known Chinese-American artificial intelligence scholar, once bluntly stated that embodied intelligence will become one of the next "Polaris issues" in the field of AI. Nvidia CEO Jensen Huang also expressed a similar view, saying that "the next wave of AI will be embodied intelligence."
Just in July this year, the project team led by Li Feifei released the latest "embodied intelligence" achievement-VoxPoser.
The principle is that the user is given natural language instructions to execute specific needs, as well as environmental information (RGB-D images collected through the camera), and then the LLM (Large Language Model) will generate the same text as the VLM (Large Language Model) based on the above information. Visual Language Model, visual-language model) interactive code.
Based on this series of operations, the system will automatically generate the corresponding "3D Value Map" for the robot to perform spatial perception. It "tells" the robot where it is and where its target is.
The biggest highlight of this achievement is that the entire robot training process does not require any additional data feeding and pre-training. In other words, problems such as predefined training and large-scale data missing that are currently required for robot training do not exist here.
It is precisely because of the above capabilities that robots can basically complete any given task as long as they comply with the relevant processes.
Regarding the achievement of VoxPoser, it can be said that it is a new milestone in the integration of the fields of artificial intelligence and robotics!
In terms of robot form, the Zhiyuan team also recognizes that humanoid robots are the ultimate goal.
The Musk we are familiar with is a loyal fan of humanoid robots.
At the recent second-quarter results briefing, Tesla CEO Elon Musk announced the latest progress in the much-anticipated Tesla humanoid robot Optimus. Musk said Tesla has produced 10 humanoid robots. Walking tests are expected to be carried out in November this year, and practical tests are planned to be carried out at the Tesla factory next year.
In addition to Tesla, Xiaomi has also developed its first full-size humanoid bionic robot "CyberOne", whose Chinese name is "Tieda". The height of the whole machine is 1.77 meters and the weight is 52KG, which is about the same size as a normal adult.
According to Lei Jun, "Tieda" has the same high IQ as humans. It can perceive 45 kinds of human semantic emotions and distinguish 85 kinds of environmental semantics. "Iron Big" adopts Xiaomi's self-developed whole-body control algorithm, with 21 joint degrees of freedom for coordinated movement, 5 joint drives throughout the body, and a peak torque of up to 300Nm; in terms of vision, through the blessing of the Mi Sense visual space system, "Iron Big" Have the ability to reconstruct the real world in three dimensions.
At the WRC 2023 not long ago, the well-known robotics company UBTECH released an intelligent multi-modal interaction system for humanoid robots. This is an intelligent system that integrates human-computer interaction, command understanding, autonomous planning, task execution and other technologies.
UBTECH also displayed the panda robots Youyou and Walker X at its booth. They are both large humanoid robots that have ridden balance bikes and just rushed to Beijing from the closing ceremony of the Universiade in Chengdu.
Overall, humanoid robots have come to the fore. In the future, embodied intelligence + humanoid will become a big killer.
write at the end
At the end of the press conference, Zhihui Jun showed the dynamic team of Zhiyuan through a video~
Starting a business is difficult, but passion seems to overcome all obstacles.
When talking about his dream, Zhihui Jun said:
“One of my dreams is to one day make the intelligent robots in science fiction movies a reality. These robots are no longer simple mechanical devices, but partners with independent thinking and learning abilities, capable of sensing and understanding. our world and then engage in deep communication and collaboration with us.”
Tribute to every dream chaser
I hope that there will be more and more people in China who are willing to challenge innovation like Zhihui Jun to create China’s own high-tech products. We also look forward to more surprises from expedition AI in the future~
Recommended in the past
Sand is made into chips, and chips are sold into sand
They are all 32-bit MCUs. What are the differences between ESP32, GD32 and STM32?
Beautiful pictures to share, the bricklayer on the circuit board