Ahead of the cold wave, a new policy specifically for humanoid robots has been released. On November 2, the Ministry of Industry and Information Technology issued the "Guiding Opinions on the Innovative Development of Humanoid Robots" (hereinafter referred to as the "Opinions").
The document outlines a promising future for humanoid robots and even provides a specific timeline. It proposes that by 2025, a preliminary innovation system for humanoid robots will be established, and mass production will be achieved; by 2027, their overall strength will reach world-class levels.
The humanoid robot field is now much more bustling and noisy. But in the eyes of outsiders, the perceptions and logic honed over many years by those well-known practitioners who have been exploring this field for so long remain shrouded in mystery, like a fog.
Therefore, I have compiled the speeches and dialogues of experts, scholars and entrepreneurs at the Intelligent Future Humanoid Robot Technology and Industry Development Forum during the World Robot Conference more than two months ago, and sorted out the thoughts and ideas of the authoritative figures in the field for reference.
01
The biggest variable in the future?
Huang Xiaoqing has been running CloudMinds for eight years, and he has a deep understanding of why robots must be humanoid, a question that has been debated in the robotics industry.
Huang Xiaoqing is the founder and CEO of CloudMinds Technology and the vice chairman of the China Cloud Computing Professional Committee. He likes to find answers in the past.
More than a hundred years ago, Ford dreamed of building a "horse-drawn carriage," and from the Model A to the Model T, the world's automotive industrial revolution began.
Humanoid robots have existed for over a century since their inception. During this evolution, there have been ebbs and flows, but the process has never ceased.
Huang Xiaoqing observed that since Japan created its first humanoid robot in 1973, its research has never stopped, even after the Fukushima incident when it entered a period of decline; Germany, another "industrial powerhouse," has also been pursuing research on humanoid robots with great dedication.
"Humanoid robots will be the first choice for human-machine coexistence in the future," Huang Xiaoqing predicted. He believes that the development and evolution of robots is similar to the transformation of AI from specialized small models to general-purpose large models, following a path from specialized robots to general-purpose robots. Under this premise, humanoid robots simultaneously possess three characteristics: automatically adapting to human environments, using human tools, and engaging in natural language interaction.
Andrew Yao, winner of the 2000 Turing Award, academician of the Chinese Academy of Sciences, and director of the Institute for Interdisciplinary Information Sciences at Tsinghua University, has recently been pondering: What is the next step in artificial intelligence?
“The future AGI (Artificial General Intelligence) will need to have a tangible entity,” Yao Qizhi concluded after careful consideration.
Yao Qizhi said that humans themselves are a rather ideal embodied intelligent entity in the world, which basically has three components: body, cerebellum, and brain. The body part should be humanoid robots.
Dennis Hong is a professor at UCLA and the founder of RoMeLa Labs. He has been fascinated by robots since childhood and continued his childhood dream as an adult.
Dennis Hong dreamed of living with robots since he was a child. He observed that the home environment is designed for humans; for example, stairs and doorknobs are used by humans. Therefore, he believed that robots should ideally be humanoid. "For the past twenty years, we have focused on research into humanoid robots," said Dennis Hong.
Marc Raibert, founder and CEO of Boston Dynamics, is a renowned long-term thinker. He believes that robotic artificial intelligence consists of two parts: motor intelligence and cognitive intelligence. Cognitive intelligence is the imagery in the human mind. Motor intelligence concerns how one uses their body.
Mark Raibert discovered that humans possess remarkably advanced motor intelligence, even children as young as a few years old. Therefore, he attempted to endow robots with motor intelligence.
Wang Xingxing, founder and CEO of Unitree Technology, admitted, "A few years ago, I was a bit pessimistic. I even felt that the world's technology had stagnated and the dividends had been exhausted." However, since the end of last year, he has rekindled his fighting spirit.
This stems from the progress in general-purpose AI, which far exceeded his personal expectations. He felt that the AI needed for general-purpose humanoid robots is gradually maturing. At the end of last year, they initiated a project to develop a general-purpose humanoid robot.
“A new industrial revolution and a new economic growth will emerge in the next five to ten years,” Wang Xingxing said.
Wang Xingxing believes that the combination of humanoid robots and general AI is the biggest variable for the future technological and economic growth of all mankind.
02
The Ambition of the Steadfast
Zhou Jian said that he mentioned many years ago that humans need humanoid robots to empower their lives.
Zhou Jian is the founder, chairman and CEO of UBTECH Robotics, a humanoid robot company he created, in which he has been immersed for 11 years.
In Zhou Jian's view, humanoid robots are not lacking in application scenarios.
In industrial settings, he observed that even with ABB and KUKA utilizing industrial robotic arms on the final assembly line, a large number of workers still operate on-site. Many leading companies employ hundreds of thousands of production line workers. This future could potentially be replaced by humanoid robots.
"Humanoid robots and traditional intelligent equipment working together to solve complex scenarios in flexible, unmanned factories—that's the future of super unmanned factories powered by robots or intelligent vehicles," Zhou Jian said.
How big is the future market potential? Zhou Jian estimates that if a car factory with 500,000 workers can reduce its workforce by 10%, that would be 50,000 workers. Assuming a saving of $60,000 to $70,000 per worker, this means that a B2B factory could have a market worth billions of dollars.
In everyday life, everything must conform to ergonomics, making humanoid robots particularly suitable for accompanying humans in this environment, working, learning, and living together.
Especially with the continuous development of AI big data models, different universities, research institutions and companies are using AI big data models to make dreams a reality.
Wang Xingxing believes that the growing popularity of humanoid robots since last year is inseparable from Musk's personal involvement, and domestic capital has also benchmarked against international standards. The field has experienced many cycles of booms and busts. "It's hard to imagine that ten years ago, AI was a niche industry. If it weren't for the explosive growth of generative AI last year, AI might have entered a new slump. But now, the pace of AI iteration is accelerating."
Mark Raibert and his team’s research dates back 15 years, during which they developed a variety of robots, such as the Big Dog robot, the Wildcat robot, the Spot robot, the Stretch logistics robot, the Atlas humanoid robot, and so on.
Mark Raibert represented Boston Dynamics' present, future, and tomorrow with three robots. The humanoid robot Atlas, currently under development, represents the future. The Spot robot and the Stretch logistics robot represent today and tomorrow, respectively.
“The humanoid robot Atlas, which we call the robot of the future, is not yet a robot in the true sense of the word,” Mark Raibert said.
The Atlas robot has an onboard computer and batteries, and approximately 28 different sensors or a vision system on its limbs to help it see the terrain ahead. Mark Raibert revealed that they are currently researching Atlas's mobility and maneuverability.
Mark Raibert believes that hardware engineering is crucial for the future of robotics, just as important as software. "Some people think that software can overcome all the problems and limitations of hardware, but I personally disagree. I believe that only through the close collaboration of the best hardware and software designers can the best robots in the world be designed."
Besides motor intelligence, cognitive intelligence is equally important. Mark Raibert stated that they have established a new institution called the Institute for Artificial Intelligence, which, with the support of the Hyundai Group, is dedicated to developing intelligent robots with both motor and cognitive intelligence.
Dennis Hong recounted the twists and turns in the development of the "ARTEMIS" humanoid robot. It took the team six years to develop. In 2018, it was merely a simple sketch on paper. In 2019, they raised $120,000 through crowdfunding to launch the project. In 2020, the pandemic hit, and the lab was forced to close just as the project was beginning. "But we didn't stop there. We continued our research at home, holding online meetings in our dorm rooms and bedrooms..." Dennis Hong recalled.
Dennis Hong further recalled, "In January 2021, it was in the dynamic evaluation period; in February, we got it to stand up for the first time, but it didn't seem very successful; in March, we started to get it to do balance work for the first time, pushing and pulling it, making the ground vibrate, and conducting various experiments such as standing and balancing... In October 2022, it walked steadily for the first time, which was a very significant success; in December, it experienced its first serious malfunction. For a force-controlled robot, if there is a problem, it will lose its balance and almost self-destruct. It was a big blow to us at the time, and it took a long time to repair the prototype afterwards."
In March of this year, the "ARTEMIS" humanoid robot was officially released. Its notable feature is its ability to maintain stability even under extreme kicking and striking tests, achieving a running speed of 2.1 meters per second. It can play soccer and even participated in the 2023 RoboCup soccer competition. Dennis Hong said that after the competition, he discovered that Bill Gates had followed him.
03
A battle between advance and resistance
In recent years, many well-known manufacturers have been forced to abandon or suspend years of research and development achievements. For example, in March last year, Honda's Asimo officially announced its retirement. It was created after 14 years of technological breakthroughs, but its fate was that after 12 years and seven iterations, it finally ended its life cycle. Before that, SoftBank's Pepper robot had already ceased production in 2021, surviving for only 6 years.
The reasons for this are twofold: firstly, the functionality did not meet the requirements that customers would be willing to pay for; and secondly, the cost was too high.
In the early days, even the most basic integrated joints required the establishment of specialized companies. However, in recent years, the rapid development of intelligent manufacturing and industrial robots has greatly integrated the upstream and downstream industrial chains, making it possible to reduce costs through increased production volume a reality.
In addition, the generative AI boom sparked by ChatGPT has opened a new door for it.
Wang Xingxing is a newcomer. He saw that existing large-scale model technology can be used directly, either by transplanting or cutting it down. For humanoid robots, a small or medium-sized model is sufficient.
Hiroshi Ishiguro, a professor at Osaka University in Japan, envisions a future where humans utilize various robots and avatars to assist in diverse activities. This implies a desire to develop a society where humans and robots coexist, rather than a society solely supported by robots.
For over 20 years since launching his interactive robot project, Hiroshi Ishiguro has created numerous robots, humanoid robots, avatars, and computer-generated avatars.
Hiroshi Ishiguro's observation is that, thanks to the advent of large language models, they can focus on studying the higher-level language cognition functions of robots.
He further stated that by integrating large language models with technologies such as facial expressions and gestures, they can study the integration of many modalities and multimodal expression, and then imbue robots or humanoid robots with intentions or desires.
“I believe that consciousness is one of the most important issues in the next 10 years,” Hiroshi Ishiguro said.
Huang Xiaoqing believes that robots are humanity's third computer, and that programming robots in the future will not be fundamentally different from programming Android phones or Windows PCs. "The emergence of ChatGPT and the development of 5G have ushered in an era of using artificial intelligence and cloud robotics technology to liberate productivity. This is a very important industrial revolution."
As the advance progresses, the resistance becomes obvious.
Huang Xiaoqing observed that the current large-scale models lack robot data; without sufficient behavioral data, it's impossible to support robot control. More behavioral data needs to be incorporated for secondary training. Once the robot possesses behavioral and spatial capabilities, the multimodal large-scale model, which could be called Robot GPT, can be further accelerated.
“Robot GPT is the brain of the robot,” Huang Xiaoqing said.
Ren Lei, a tenured professor at the University of Manchester and the Tang Aoqing Chair Professor at Jilin University, has been working on two things simultaneously for more than 20 years: bionic robots and human science.
Ren Lei's deep feeling is that, compared with humans, humanoid robots are far behind in terms of mobility, including their mobility economy, mobility agility, mobility diversity, environmental adaptability, and mobility plasticity.
"Now we've turned our attention back to the human body. How is the human body designed? What insights can it offer? Can we develop a new biomimetic technology that can lead to new types of humanoid robots? These are the questions our team is currently exploring," Ren Lei said.
Professor Alois C. Knoll of the Technical University of Munich believes that the main obstacle to humanoid robots currently lies in their form. Building a computer for a data center is relatively easy, but creating a robotic arm that completely mimics a human arm is much more difficult, making it challenging to enable robots to walk like humans.
In addition, Alois C. Knoll said that issues such as energy supply also require breakthroughs in new technologies. There is also a more distant problem: how to enable a system to develop its own intelligence, such as its own perception or soul, which involves ethical issues.
Wang Xingxing said that the general-purpose large model itself is not intended for general-purpose robots. It has already processed the text or textual logic, but it is static and has virtually no environmental cognition or perception. Without the ability to perceive the physical environment, it cannot grasp or perform any operations.
Wang Xingxing observed that Google is very dedicated to this matter, having been working on robotics and AI for the past ten years. "Everyone is now eager to create a large, dedicated model for humanoid or general-purpose robots."
Chen Jianyu, Assistant Professor at the Institute for Interdisciplinary Information Sciences at Tsinghua University and founder of Star Era, believes that while robotics technology has made significant progress in both hardware and software, many technological bottlenecks remain. For example, in terms of hardware, robots need to balance factors such as strength, speed, precision, and cost, but currently no single technology can simultaneously meet all these requirements. Hydraulic technology offers high strength and speed but is expensive, while high-reduction-ratio harmonic motor technology offers high precision but limited flexibility. Therefore, the current solution is to balance various technologies, selecting different technological approaches based on different application scenarios. In the future, people hope for a new technology that can balance all factors like a human.
Chen Jianyu discovered that the software also has significant problems. Currently, it cannot effectively balance generalization and the precision of task operation. For example, Google's Robotic Transformer series can achieve high generalization, but its control precision is relatively low, resulting in a low control frequency. It primarily performs tasks at the operational control level, suitable for relatively simple tasks.
Chen Jianyu also observed the issue of safety. While the various problems that arose during the experiment were somewhat explainable, it was difficult to explain whether a potentially dangerous action might occur using large models and neural networks. "When using ChatGPT for dialogue, it might mispronounce some words, which isn't a big deal. However, as a robot entering our lives, this becomes crucial, and these are all areas that urgently require new technologies to improve."
Alois C. Knoll also poured cold water on the idea, stating that humanoid robots have been evolving for over a century, with very slow progress. He cautioned against setting overly high standards or creating a speculative atmosphere that could turn into a bubble. "This is extremely difficult, high-level research, and it may be slower than general AI; we may not see rapid, dramatic changes."
“There are still many mountains to climb for humanoid robots,” Alois C. Knoll emphasized.
