news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Text｜Liu Junhong

Editor: Wang Yisu

In the stuffy venue, there was excited discussion, and everyone was afraid of missing out on this feast of AI-enabled robot evolution.

Intelligent shooting of light cones at a crowded exhibition site

On August 21, the 2024 World Robot Conference (WRC) opened in Beijing. In this exhibition with 169 companies and more than 600 products, embodied intelligence almost "contracted" an entire venue, and humanoid robots and robot dogs from 27 companies attracted the most visitors.

In the demonstrations of many robots, LightCone Intelligence found that the ability of each company to perform complex tasks has been significantly enhanced and the practicality has also been greatly improved.

Compared with last year, robots demonstrated more simple tasks such as dancing, talking, and walking. This year, there are Weijing intelligent robots picking fruits, UBTECH robots repairing cars, Xingchen intelligent robots writing calligraphy, and humanoid robots playing football games. In addition to the robot dog that performed continuous sideways flips and jumps, Yusu Technology also brought a humanoid robot, the G1, priced at 99,000 yuan. With the demonstration of the capabilities of robots and the release of prices by many manufacturers, humanoid robots are getting closer and closer to becoming a reality.

Yushu Technology G1 robot action demonstration

“The development of large models has greatly promoted the progress of humanoid robots.”

Liu Cong, vice president of iFLYTEK and director of the research institute, told Light Cone Intelligence that both the "brain"'s ability to perceive and understand and the "cerebellum"'s ability to control movement have been greatly improved.

Furthermore, Zhang Li, co-founder and COO of Zhuji Power, believes that the evolution of AI has enabled robots to achieve a leap forward. "General artificial intelligence enables software and algorithms to help robots achieve many things that were previously impossible to accomplish."

With the breakthrough of AI big models, today's humanoid robots are beginning to develop in the direction of autonomous driving. In the evolution of the end-to-end big model that allows cars to understand the world while iterating rapidly, robots have also demonstrated the ability to understand complex instructions and learn smoother movements.At the same time, under the demand for large model training, humanoid robots have also entered the stage of data-driven intelligent iteration. Robot manufacturers are also trying to gain a foothold in a series of scenarios such as home, industry, warehousing and logistics, and retail. They hope to achieve a closed loop of data and iteration centered on a single scenario, connect different fields, and eventually move towards universal use.

Although robots have experienced many hopes and disillusionments in the past decade of AI development, Stardust Intelligence founder Lai Jie excitedly said, "In the next decade, the most worthwhile thing to do is humanoid robots" based on the clear evolutionary direction of robots this time.

Robots are setting off a carnival of capital, technology and industry.

Capital believes in humanoid robots

"In the past two years, the most investment we have seen, besides large models, is in robots," Liu Cong told Lightcone Intelligence.

Although the humanoid robot industry is still in the stage of capability demonstration, investors and entrepreneurs know that they cannot miss this opportunity.

After reviewing the investment trends over the past year,We found that humanoid robots, as the ultimate form of robots and the hottest type, have the highest technology and difficulty, and have also occupied a high ground in the capital market.

At the entrepreneur level, ITjuzi data shows that since January 2023, a total of 29 humanoid robot companies have been established in China, of which 22 have received at least one round of financing. Among them, Zhiyuan Robotics, founded by former Huawei genius "Zhihuijun", completed seven rounds of financing within 17 months of its establishment, and its pre-investment valuation has reached 7 billion yuan.

From the perspective of capital, according to incomplete statistics from China Electronics News, there were more than 22 financing events in the field of humanoid robots in the first half of 2024, with a financing amount of more than 7 billion yuan. With the enthusiasm for investing in humanoid robots, many investment institutions have made layouts that can be described as "sweeping the market".

Cao Wei, partner of BlueRun Ventures, said that he has invested in more than 10 early-stage projects in the field of robotics. Matrix Partners also owns four humanoid robot companies, namely, Yushu Technology, Zhiyuan Robotics, Galaxy General, and Stardust Intelligence.

"The industry and capital are very optimistic about the future of robots. After all, this is a trillion-dollar market," UBTECH's global marketing director Li Zhuo told Cone Intelligence at the conference.

Therefore, under the combined effects of software, hardware, capital promotion, and cutting-edge application effects, humanoid robots have become the hottest track in the AI ​​era.

Take the experience of Yushu Technology, which only started exploring humanoid robots in early 2023, as an example. Wang Xingxing believes that the reason for not making humanoid robots before is that "the control technology of humanoid robots in the world is not very ideal. The performance is not up to standard and cannot reach the stage of practicality or work, so humanoid robots have not been made for many years."

Today, with the upgrading of robot software and hardware technology, humanoid robots are gradually moving from laboratories to actual application scenarios. In July this year, Musk said that two Optimus robots were already in Tesla's factory carrying batteries. At this robot conference, we also saw UBTECH's humanoid robots being able to perform simple automotive quality inspection tasks.

UBTECH Robot Demonstrates Quality Inspection Scene

Admittedly, measured from the perspective of skilled human workers, the efficiency of robot "employees" is still too low. However, if combined with the scale, the 24-hour working characteristics of humanoid robots have attracted the willingness of different industries such as automobiles, logistics, scientific research institutes, and AI technology to try, and jointly urged the delivery process of humanoid robots.

"At the end of 2022, we have not yet made humanoid robots, but some customers have already approached us to buy humanoid robots," Wang Xingxing believes that Yushu Technology's progress in humanoid robots is the result of going with the trend.

However, Fu Sheng, chairman and CEO of Cheetah Mobile and chairman of Orion Star, believes that there are still many aspects of the robotics industry that need to be polished, especially humanoid robots that "walk on two legs" face the limitations of physics. "The technology of bipedal robots is too complicated, and it must rely on mechanical structures. The iteration of mechanical structures will not be as fast as autonomous driving, and the progress each year will not be much."

But in short, industry players have reached a basic consensus on the expected maturity time of humanoid robots. Li Zhuo told Lightcone Intelligence,"It is estimated that the industry will take another 3-5 years to fully mature."

In the eyes of leading entrepreneurs, humanoid robots at this moment are just like in 2019, when Tesla had just launched the FSD chip and Baidu Apollo had just obtained multiple test licenses for autonomous driving. Everything is full of hope.

All this is because humanoid robots have undergone fundamental changes under the influence of this round of AI big models.

AI evolves, robots become more human-like

“One is a large language model, and the other is an end-to-end algorithm.”

Xie Chen, founder and CEO of Guanglun Intelligence, summarized the biggest reason for this round of robot evolution.

At the conference, humanoid robots, robot dogs, and even catering robots with differentiated designs all showed "more reasonable" movements. Compared with the one-year cycle, the robots at this year's conference were much smoother.

A more obvious example can be seen in the simple task of making ice cream by a robot. At last year's World Robot Conference, the robot arm's movements were basically monotonous "turning in circles". This year, the ice cream robot of AUBO Intelligent has a much smoother movement, and the robot arm only turned a circle to pick up the ice cream.

Comparison of the movements of the ice cream robot in 2023 (top) and 2024 (bottom)

Different degrees of movement smoothness mean that the underlying logic of the robot's task execution has changed.

"The original robot operation design was modular, with the underlying solution being a mixture of learning and rules," said Zhao Xing, co-founder of Xinghaitu, assistant professor at the School of Interdisciplinary Information Sciences at Tsinghua University, and director of the MARS Lab."This model is a bit like the object detection-decision-planning-control process of previous autonomous driving. For example, when making a grasping strategy, we need to detect the object, estimate its state and posture. But most objects in reality have no posture at all, just like a piece of paper spread on a table. It is difficult to define the posture and key points."

This is exactly the same as the development process of autonomous driving. Before the development of robot motion planning and autonomous driving to "end-to-end", they encountered the same problem:In the development mode of specific scenarios, only limited actions can be written based on rules, which cannot adapt to the infinite scenarios in the real world.

Robots can only operate objects of limited shapes and types, just like autonomous driving can only be used on closed roads. With "end-to-end" machine learning, just like a car can learn how to turn around by itself, robots can also learn how to grasp objects of different shapes, colors, and softness.

UBTECH robots listen to instructions and grab things

Referring to the trend of combining autonomous driving with big models, robots can also begin to understand the real world after combining with multimodal big models. At the scene, Galaxy Universal Robots demonstrated the scene of pharmacy duty. In addition to taking medicines from the shelves, the robot can also pick up things that have fallen to the ground by itself.

Galaxy Universal Robots demonstrates pharmacy duty scene

Liu Cong believes that the importance of AI big models to improving robot capabilities is mainly reflected in three aspects:

First, the large model significantly improves the robot's ability to solve complex tasks, breaking down complex tasks into a series of executable tasks based on understanding.

Next, with the capabilities of multimodal large models, robots can rely on vision, touch, etc. to perform tasks in more complex scenarios.

Finally, in terms of athletic ability, the robot can perform simulation training based on AI-generated data. Without a large model, the functions that the robot can achieve will be relatively limited.

It can also be seen from this thatIn addition to large model technology, data is the most critical factor restricting the next evolution of robots.

In order to obtain enough data for robot training, many manufacturers have solved the problem by combining real data with simulated data. In obtaining real data, manufacturers generally "feed" the data during the task to the large model through real-life actions and robot teleoperation. Simulation data is obtained by building a scene that is as real as possible, setting as many modeling and reality parameters as possible, and allowing the virtual robot to be trained.

However, the above two methods still have their own shortcomings, and the current robotics industry is far from the start of the data flywheel. On the one hand, the cost of real data is extremely high, and robots are not yet as popular as autonomous driving. Before there is enough real data, manufacturers need to spend a long time and equip data standard teams to accumulate it bit by bit. As for simulation data, the biggest problem is that the reality simulation is not real enough. Robots can succeed in the simulated world, but there will be a certain failure rate when they are introduced into the real world.

"At this stage, the biggest limitation to the robotics industry is that AI is not enough. AI models, AI training data sets, and AI scenario deployment are all far from enough," said Wang Xingxing.

With the development of AI,Zhao Xing predicts that "with more efficient algorithms and improved generalization capabilities, the amount of data required for robots to learn a skill will drop from the current thousands or tens of thousands to a thousand or even hundreds or dozens of orders of magnitude in the future."

Referring to the data flywheel brought by mass production in the autonomous driving industry, in order to further solve the data problem, many robot manufacturers are exploring how to "get" data from mass production.

Grab sales

Humanoid robots on the eve of mass production

Price is a major constraint on the mass production of humanoid robots.

According to what Lightcone Intelligence learned from various companies at the World Robot Conference, Jiang Qingsong, partner and vice president of marketing services of Zhiyuan Robotics, said that the industry price of a 1.7-meter-tall humanoid robot is about 600,000 to 700,000 yuan.

Compared with Boston Dynamics, which started at millions in previous years, these prices are already "trial prices" acceptable to a few industries, but they are still a long way from large-scale popularization.

Wang Xingxing said, "We have not yet reached a true commercial closed loop. A robot cannot achieve a lower cost than a human, and its commercial value is still not positive."

This is because, from the perspective of R&D costs, due to the immaturity of humanoid robots, the company's R&D process will set up sufficient redundant space in the links of technology layout, hardware selection, scene design, etc. This results in the product "carrying" a lot of "hidden costs" in addition to the necessary hardware under the allocation of early R&D costs.

After the sales phase is over, the robot company still has to be responsible for the application effect. "Manufacturers should assign a dedicated on-site team to carry out subsequent R&D and maintenance for industrial application pilot customers," an exhibitor told Lightcone Intelligence at the conference.

Although current customer demands and products are very non-standardized, LightCone Intelligence found at the conference that many manufacturers have tried to develop corresponding product strategies based on different demands in an attempt to make robot sales smoother.

Basically, Lightcone Intelligence learned that the prices of general humanoid robots from many manufacturers are concentrated in the range of 500,000 to 600,000 yuan, and they are supplemented with cheap mass-produced versions. For example, Yushu Technology, which has the highest sales of quadruped robots, has set the price of its general humanoid robot H1 at 500,000 to 600,000 yuan, and the latest G1 humanoid intelligent body has also released a mass-produced version, priced at only 99,000 yuan.

On the other hand, the sales of Zhongqing Robot's humanoid robots are even more extreme. In order to reduce costs, after removing the robot's "head and arms", Zhongqing DG01 simply lowered the price of the "big bipedal robot" to 38,500.

This may mean that the first round of price war for humanoid robots is coming soon.

ZhongQing Robot is priced at RMB 38,500 and can shoot with light cone intelligence

But larger-scale price cuts still depend on mass production.

Recently, Zhiyuan Robotics expects to ship 200 units this year, which is already the industry leader. In the previous stage, the shipment volume of commercial cleaning robots by a single company was only around 1,000 units a year.

When the sales volume can exceed 10,000 or even 100,000 units a year, that will be the real "iPhone moment" for robots.

In order to achieve the goal of mass production, leading humanoid robot manufacturers have already set their sights on factories as the first scenario.

Globally, in addition to Tesla's own robot Optimus, Figure, which OpenAI invested in, has cooperated with BMW and plans to deploy Figure 01 robots in South Carolina, USA. Apptronik, an American robot manufacturer, has cooperated with Mercedes-Benz, and its Apollo robot has been moving boxes on Mercedes-Benz's assembly line.

In China, the fastest-growing company is UBTECH Robotics, the only listed company that produces humanoid robots. Since the beginning of this year, UBTECH's humanoid robots have been "working" in many auto companies, including NIO, Dongfeng Liuzhou Motor, Geely Auto, and FAW. With UBTECH's deployment in classic manufacturing scenarios, humanoid robots have initially verified the feasibility of industry applications in quality inspection, moving goods, sorting, screwing, and assembly.

"The robotics industry is still in its early stages, but with the support of large models, technology and products are expected to break through from 0 to 1 within three years," Xie Chen concluded.

Source: Some pictures are from the Internet. If there is any infringement, please contact us to delete them.