The Robots Are Coming: Unitree’s Rise

Humanoid robotsleave science fiction and enter factories, research laboratories and public stages. At the center of this development is Unitree founder Wang Xingxing. His company builds inexpensive robot dogs, humanoid work machines and even a controllable mecha with the GD01. But behind the spectacular performances is early-stage technology. The machines lack general understanding, reliable operating times and sufficient training data. At the same time, competition worth billions is growing betweenChinaand the USA. “The Robots Cometh” therefore not only describes the arrival of new machines. It’s also about work, power, security and the question of how much physical agency we want to leave to artificial intelligence.

The most important thing in brief

Unitree is currently a leading manufacturer of humanoid robots in terms of unit shipments. In 2025, the company sold more than 5,500 units.
The 2.74 meter tall GD01 weighs around half a ton. It can transport a person, walk on two or four legs and is expected to cost $650,000.
Humanoid robots have so far only achieved around 30 to 50 percent of human efficiency in general factory and warehouse tasks.
The most important bottleneck is no longer just the mechanics. Versatile AI models and large amounts of high-quality movement and environmental data are missing.
Humanoids could take on dangerous and monotonous tasks. At the same time, significant risks arise for labor markets, data protection, cybersecurity and the distribution of wealth.

Unitree’s GD01 turns the mecha into a real machine

The idea for the GD01 came about after attending a UFC fight. While other spectators might want to train themselves afterwards, Wang Xingxing went back to his laboratory. At first he thought about a larger robot for boxing competitions. This led to the idea of an even larger machine that could even hold a human. The result is the GD01, which Unitree describes as the world’s first mass-produced and transformable mecha robot.

The GD01 is nine feet or around 2.74 meters high. It weighs about half a ton. In its fuselage there is a cockpit for a pilot. The machine can run on both two and four legs. Their arms and fists are said to be so strong that they can tear down walls. Resilient joints connect titanium alloy limbs with a protective carbon fiber outer skin. This means that the massive robot should move in a controlled and comparatively fluid manner.

Wang presented the GD01 exclusively for TIME in June 2026. It was also his first interview with an international media. In addition to martial arts, James Cameron’s film “Avatar” also apparently influenced him. According to Wang, the human-controlled combat robot shown there served as a visual reference. The GD01 deliberately combines well-known science fiction images with real engineering technology.

However, the presentation also shows how far apart expectations and reality still are. At Unitree’s headquarters, the GD01 hung from a reinforced steel beam. Wang didn’t want to operate it or get into the cockpit himself. He justified this briefly with possible security problems. The robot is therefore not yet a fully usable everyday product. Its listed retail price of $650,000 also makes it a machine for specialty applications, research, or wealthy technology enthusiasts.

feature	Unitree GD01
Construction type	Transformable, pilotable mecha
Height	Around 2.74 meters
Weight	About 500 kilograms
locomotion	On two or four legs
material	Titanium alloy and carbon fiber
steering	Pilot in fuselage cockpit
Stated ability	Very strong blows, transporting a person
inspiration	Martial arts and a mecha from “Avatar”
Selling price	$650,000
Level of development	Impressive prototype with open security questions

However, it is not the spectacular GD01 that is most important for the upcoming robotics change. This role is more likely to fall to the smaller G1. It can already carry limited loads and is being tested in initial logistics and material transport projects. While the GD01 attracts attention, the G1 is intended to be a practical workhorse. It is precisely this dichotomy that shapes Unitree’s strategy: spectacular machines create public visibility, while cheaper models open up new markets.

Wang Xingxing and the rise of Unitree

Wang Xingxing comes from the Chinese coastal city of Ningbo. He became interested in robotics at the age of ten. Television images of early robot experiments at the Massachusetts Institute of Technology gave him an important impetus. He was particularly influenced by the work of Marc Raibert, the later founder ofBoston Dynamics. In his free time, Wang built model airplanes, conducted technical experiments, and ultimately designed small turbojet engines.

In 2009, he began studying mechatronics at Zhejiang Sci-Tech University in Hangzhou. According to his own statements, he built a small two-legged robot as a freshman. The device could walk a few steps and cost less than $30. This early focus on low costs later became a core element of Unitree. Wang wanted to make robotics not only more powerful, but above all more affordable.

During his master’s degree at Shanghai University, he developed the four-legged XDog. The prototype is still visible today at Unitree headquarters. Unlike the hydraulic robots from Boston Dynamics known at the time, XDog worked with electric motors. This reduced effort and costs. At the same time, the project won second prize in a national start-up competition.

After graduating, Wang worked briefly for drone manufacturer DJI in Shenzhen. He then founded Unitree in 2016 at the age of 26. An angel investor provided two million yuan for this. That was equivalent to about $300,000 at the time. Within a decade, the young company became a globally respected robotics manufacturer.

station	Development
childhood	Enthusiasm for MIT robotics and Marc Raibert
Leisure projects	Model airplanes, experiments and mini turbojets
2009	Start of studying mechatronics in Hangzhou
First year of study	Bipedal robot for less than $30
Master’s degree	Development of the electrically powered XDog
Career entry	Worked at drone manufacturer DJI
2016	Established Unitree with two million yuan
Further development	Robot dogs, humanoid models and the GD01

Wang’s public image differs from many prominent technology founders. He appears reserved and largely avoids ostentatious self-dramatization. Nevertheless, he speaks with great conviction about the long-term importance of humanoid machines. In ten years, he believes every household could have a small robot. This could do laundry, prepare food or provide simple help with health problems. Dangerous work on mountain slopes or in sewage systems would also be possible areas of application.

Unitree has also become a Chinese cultural phenomenon. An estimated 600 million people saw footage of the H2 and G1 models at China’s Spring Festival in 2026. The robots wielded nunchucks and performed flips and choreographed martial arts moves. It was the second year that Unitree robots were among the most important program items at the gala. According to Wang, he did not yet know what the company would show at the next edition.

However, popularity creates significant pressure. In March 2026, Unitree filed documents for an IPO. The target valuation was six billion US dollars. As a result, share prices of possible suppliers also rose on Chinese stock exchanges. At the same time, the public and investors constantly expect new technological advances. Wang therefore warns that real advances in deep technology take time.

In the first quarter of 2026, Unitree’s revenue increased more than 68 percent year-over-year to $62 million. However, adjusted net income fell by half to $6 million. Unitree attributed this to significantly higher research, development and sales costs. Growth and profitability do not automatically develop at the same pace. The company must meet high expectations while solving fundamental problems in robotics.

China is leading the global race for humanoid robots

Humanoid robots are considered the next big technology platform. They combine artificial intelligence with a body that can move in an environment designed for humans. Such a machine should see, hear and understand language. She is also said to grasp objects and physically change her surroundings. This turns AI from a digital tool into an active actor in the real world.

Morgan Stanley expects up to 13 million humanoid robots worldwide by 2035. This roughly corresponds to the current population of Bolivia. By 2050, the inventory could rise to one billion units. The entire market could then reach a value of five trillion US dollars. However, these are long-term forecasts. They require strong technical advances and a high level of social acceptance.

Unitree was a market leader by unit volume in 2025. The company delivered more than 5,500 humanoid robots. This meant that the Chinese manufacturer accounted for over a quarter of the still small global market. In June 2026, Nvidia also announced that it would use its powerful Jetson Thor chip in Unitrees H2. The model is intended to be sold to research institutions in various countries.

The competition is still growing quickly. Tesla introduced the third generation of its Optimus in January 2026. In the long term, the group aims to produce ten million units per year. They will be manufactured in Fremont and Texas. However, this target is far higher than the current sales figures for the entire industry.

The Californian company Figure sent its model at the end of JuneFigure 03to the BMW plant in Spartanburg in South Carolina. Its predecessor, Figure 02, was already involved in the assembly of 30,000 vehicles there. Agility Robotics from Oregon, in turn, reported operations at Toyota and Mercado Libre. This creates the first practical applications in industry and logistics. At the same time, it remains unclear how independently and economically the robots used actually work.

Pursue	Robot or project	Stated level of development
Unitree	G1, H2, R1 and GD01	Market leader by number of units in 2025
Tesla	Optimus Generation 3	Long-term goal of ten million units annually
figure	Figure 02 and Figure 03	Testing and use in the BMW Spartanburg plant
Agility Robotics	Digit and other systems	Projects at Toyota and Mercado Libre
UBtech	U1	Humanoid companion robot with biomimetic face
Noetix	Humanoid platforms	Research on manipulation and generalization

China considers this industry to be strategically important. President Xi Jinping calls AI and robotics important new growth drivers. The Chinese Ministry of Industry and Information Technology classifies humanoids as a disruptive innovation. Their importance should be comparable to smartphones and electric cars. Technological leadership is therefore seen as an economic and security policy task.

Since the end of 2024, Chinese cities have raised robotics and technology funds totaling more than $26 billion, according to Morgan Stanley. Unitree does not disclose how much government support the company will receive immediately. However, as one of Hangzhou’s “Six Little Dragons,” the company benefits from the local sci-tech fund worth $14 billion. In spring 2025, Wang also became the youngest participant to take part in a high-level economic symposium with Xi.

In China, more than 140 companies are now working on humanoid robots. Unitree is not just another manufacturer there. Observers see this as a possible hardware platform along the lines of BYD for electric cars or DJI for drones. Elon Musk also appreciated China’s position unusually clearly. In a Tesla conversation in January 2026, he said that there were no significant competitors for humanoid robots outside of China.

The rivalry is already having political consequences. Three U.S. Representatives introduced the bipartisan Guard Act on June 3, 2026. He should be able to ban Chinese robots if they are considered a threat to national security. Republican Rep. John Moolenaar specifically named Unitree. He accused the company of generous government funding and possible crowding out of American competitors. The robot race therefore affects not only products, but also supply chains, raw materials, data and geopolitical power.

Low prices meet limited practicality

Unitree’s business model is based on low production costs and rapidly falling sales prices. The company develops important core components itself. These are particularly relevantActuators, the muscles of a robot. They convert electrical energy into controlled movements. Depending on the design, these components account for 50 to 70 percent of the cost of a humanoid.

The net price of the G1 fell from $16,000 to $13,500 within 18 months. SemiAnalysis estimates its components cost less than $9,000. This would result in a gross margin of approximately 33 percent. The smaller R1 model is already available for less than $5,000. Unitree’s four-legged robots also became significantly cheaper. Within six years, their price dropped from around $45,000 to less than $2,000.

Unitree product	Said price or value	Classification
GD01	$650,000	Pilotable mecha and special machine
G1, earlier	$16,000 net	Price before reduction
G1, current	$13,500 net	Flagship for research and initial pilot projects
G1 components	Under $9,000	Estimation by SemiAnalysis
R1	Under $5,000	Cheap entry into humanoid robotics
Robot dogs, back in the day	Around $45,000	Price about six years ago
Robot dogs, current	Under $2,000	Result of the cost and scaling strategy

The G1 does not achieve the top values of western competitors in all areas. This primarily concerns load capacity and durability. However, according to the report, it has exceeded a threshold of sufficient practical suitability. It can carry a load of five kilograms. You can do this for ten to 15 minutes at a time before overheating becomes a problem. Therefore, the G1 is being tested for the first time outside of pure development laboratories in logistics and material transport projects.

Unitree is still a long way from being a real mass market. Current humanoid robots only achieve about 30 to 50 percent of human efficiency in general warehouse and factory tasks. This includes carrying, stacking and sorting boxes. The purchase price alone does not determine the economic viability. Companies also need to consider maintenance, downtime, monitoring and slow workflows.

The distribution of sales shows the immature market particularly clearly. Only nine percent of Unitree humanoids go directly into industrial applications. Another 17 percent is used in tourist facilities, exhibitions and shops. The remaining 74 percent buy universities, research institutions and individual developers. These groups are primarily working on better AI, perception and motion control.

Area of application	Share of Unitree’s humanoid sales
Universities, research and developers	74%
Tourism, exhibitions and retail	17%
Industrial applications	9%

The industry primarily sells robots to people who want to develop better robots. Grace Shao, editor of the newsletter AI Proem, therefore believes that humanoids are not yet ready for the mass market. Wang, on the other hand, compares the situation to the early years of the personal computer. PCs were also initially purchased primarily by professionals and companies. With better software and a growing developer community, the proportion of real industrial applications is expected to increase quickly.

The discrepancy between performance and work became visible during the recordings in Hangzhou. An advertising background could not be moved because it would have influenced the lidar sensors of a group of training G1 robots. After just a four-minute martial arts demonstration, the sensors were so hot that technicians sprayed them with coolant. The machines can show impressive choreographies. But even small changes in the environment and short periods of stress continue to lead to problems.

The generalization gap determines the breakthrough

The biggest technical hurdle is the so-called generalization gap. A robot today can learn a clearly defined task. However, it is difficult for him to transfer this ability to a changed situation. Even a moved cup, different light or a new table surface can significantly reduce the success rate. Wang calls this problem one of the biggest challenges in global research.

Spectacular backflips should therefore not be confused with general intelligence. Such movements are precisely rehearsed and technically very limited. However, if a robot is to pick up a pen and place it somewhere else, numerous individual steps often have to be programmed. An everyday humanoid would have to understand more abstract instructions. Finally, the “go to the supermarket” command includes navigation, planning, object recognition, communication, and dealing with unexpected events.

Wang expects the defining “ChatGPT moment” for humanoids to occur in two to ten years. His definition for this is comparatively concrete. A robot would have to be able to carry out 80 percent of spoken commands in an environment that is 80 percent unknown. If that threshold is crossed, large commercial rollouts could begin. However, the broad time span shows how uncertain developments remain.

Two problems reinforce each other. First, large amounts of high-quality spatial data are missing. Secondly, real locations differ significantly from one another. Large language models can use text, images and other content from large parts of the Internet. Robots, on the other hand, need data about grip, force, movement, balance and the spatial consequences of an action.

Such manipulation data can only be collected with great effort. People or teleoperated robots have to demonstrate movements repeatedly. This includes folding shirts, stocking a fridge or making a cocktail. Each task includes numerous variants. Other objects, lighting conditions and materials can render an already trained behavior unusable.

Jiang Zheyuan, CEO of Beijing robotics company Noetix, describes exactly this problem. Even a well-trained movement strategy can lose a lot of reliability with small changes. Therefore, it is not enough to train a machine in a controlled laboratory. She has to gain experience from many homes, factories, shops and outdoor areas.

In New York, Micro AGI is taking an unusual approach. The company, which cooperates with Unitree, has employees equipped with cameras clean apartments free of charge. This creates spatial training data for physical AI. The free service is therefore not a classic cleaning offer. Its main purpose is to build a valuable data set about household work.

This shifts the actual competition away fromHardwareto the robot brain.Vision-Language-Action Models, shortVLA, are intended to connect visual perception, language and actions. A VLA model must recognize what a person wants. It then has to assess the situation and create a safe sequence of movements. Kaiso Research expects this submarket to reach a volume of $40.50 billion by 2035.

Several large technology companies are already working on such systems. Nvidia develops GR00T N1.7.Google DeepMindworks with RT-2-X. Alibaba is pursuing the Qwen Robot Suite. The future market leader does not necessarily have to be the manufacturer with the best mechanics. The deciding factor could be who has the best data, models and learning networks.

Unitree has a disadvantage compared to Tesla in this competition,Xiaomiand XPeng. These corporations have additional revenue streams, large supply chains and numerous connected products. Unitree’s absolute sales are still small despite its current leadership position. A competitor could therefore quickly overtake you if it closes the generalization gap earlier. Current market leadership is therefore no guarantee of long-term dominance.

6G, the world of work and security determine the future of robots

6G could fundamentally change the development of humanoid robots. Future AI-native networks are expected to connect millions of devices in near real time. If a robot in a warehouse learns how to use an unfamiliar tool, it could immediately share this experience with an entire fleet. This means that each machine would not have to relearn all the tasks independently. Qualcomm boss Cristiano Amon even associates this with the vision of a digital twin of the world.

Fast networks could also ease two other bottlenecks. Robots must simultaneously process 4K video images, lidar data, sound, and tactile and body signals. The computing technology required for this increases weight and power consumption. This limits mobility andBattery life. Part of the calculation could be outsourced to the cloud in the future.

The report mentions a possible latency of 0.1 to one millisecond for 6G. Under ideal conditions, cloud-based AI could then act almost in real time. The physical robot would become lighter while its powerful brain works in the network. In the event of an error, a human could also intervene remotely. This teleoperation would be an additional level of security.

However, such centralization creates new areas of attack. Cybersecurity professionals found a backdoor in Unitrees in 2025Robot dogGo1. Third parties could register via a portal. They could see the location, get live images and even control the robot remotely. Unitree then switched off the affected cloud service.

Wang emphasizes that Unitree’s robots can operate completely without a network connection. They should not store any environmental data without the customer’s consent. He also considers the amount of data stored locally to be small compared to a smartphone. However, this reasoning does not answer all security questions. Particularly with cloud-based control, access protection, logging, updates and clear emergency functions would be indispensable.

The main difference to conventional AI lies in the physical effect. A language model can produce false or dangerous information. A humanoid robot can immediately translate an error into a movement. This means that a cyber attack moves from the screen into homes, kindergartens, streets and factories. Even today’s demonstrations already show physical risks. In videos, Unitree robots’ flailing arms or legs hit children who unexpectedly came within their range of motion.

Wang still believes that out-of-control robots are unlikely. It refers to fixed hardware limits and multi-level protection systems. However, as your independence increases, checking such statements becomes more important. A system that is supposed to handle new situations independently cannot be secured solely through rigid movement rules. Security must therefore be designed mechanically, software-wise and organizationally at the same time.

The world of work is also facing profound change. Humanoids could work on production lines around the clock. They could harvest fruit in high heat, maintain dangerous canals or relieve medical staff of simple care tasks. This would free people from stressful activities. At the same time, many manual and service-oriented professions could come under pressure.

In the optimistic scenario, production increases sharply while costs decrease. This could lead to greater prosperity and a world with significantly less material scarcity. In the negative scenario, robots, data and means of production are concentrated with a few owners. Income from human labor would then become less important. States would have to adapt their social security systems while many people struggle for professional roles and social recognition.

Added to this is the competition between China and the USA. Both countries lead in both AI and humanoid robotics. They compete for raw materials, chips, data, skilled workers and industrial standards. Security laws like the proposed Guard Act could divide markets into geopolitical blocs. Companies would then have to select their robots not only based on performance, but also based on origin and data architecture.

The private household represents another hurdle. UBtech presented the humanoid companion robot U1 on June 30, 2026. He has silicone skin and biomimetic facial expressions. A supposedly empathic language model is said to recognize more than 20 finely graded emotional states. This brings robotics closer to care, support and emotional interaction.

However, technical progress does not guarantee acceptance. The Japanese roboticist Masahiro Mori described the “Uncanny Valley” in 1970. Accordingly, human-like machines initially appear more familiar the more realistic they become. At a certain point, however, sympathy turns into discomfort or disgust. Silicone skin, real hair and almost human facial expressions could increase this reaction. ForHousehold robotsSocial design is therefore just as important as mechanics and AI.

New perspective:The often overlooked bottleneck in the future may not be the number of robots, but the infrastructure around their errors. Millions of humanoids need maintenance services, spare parts, insurance, accident rules, digital identities and traceable decision protocols. It must also be clarified who is liable in the event of a cloud-supported wrongdoing: manufacturer, AI provider, network operator, owner or the person controlling remotely. The actual mass market therefore does not start with an impressive robot. It only begins when companies can deal with an ordinary defect safely, affordably and legally.

Unitree’s practical demonstrations highlight the current contradiction. The GD01 impresses with its size and power, but was not demonstrated for safety reasons. The G1 models master precise martial arts, but react sensitively to changes in their sensor environment. After four minutes, their heated lidar systems needed to be cooled. The robots therefore already embody intelligence without possessing it in a general form. They play the role of future everyday machines, but have not yet permanently settled into this role.

Conclusion

“The Robots Cometh” shows a future that has already begun, but is still full of contradictions. Unitree lowers prices, increases volumes and turns science fiction into moving hardware. Nevertheless, today’s humanoids remain slow, prone to disruption and highly specialized. The decisive breakthrough depends on data, VLA models, secure networking and social rules. If he succeeds, robots could take on dangerous work and create wealth. If the control fails, there is a risk of concentration of power, job losses and physically effective cyber attacks. The robots are coming – but no one yet knows under whose control.

Source:https://time.com/article/2026/07/23/unitree-china-human-robotics/

Bewerte den Beitrag hier!

[Total: 0 Average: 0]

Nico Nuss

Website

Author Nico Nuss has been working on mobile computing and automation software since 2001. Drawing on his experience and strong interest in future technologies, he focuses on robotics and AI.