Are the three giants of Silicon Valley igniting a mass production revolution, and will China's embodied intelligence claim the C position on the global stage?

Lead-in: Preset actions are today’s entry ticket, and generalization ability is tomorrow’s championship ticket.

Edited by | Jingcheng

Author | Jiang Jing

At the close of the first quarter of 2026, a global, synchronized action in the tech industry has officially announced a historic turning point for the humanoid robotics industry.

The three Silicon Valley giants—Google, Amazon, and Tesla—stepped up in the same week. From technology enablement and scenario planning to mass-production rollout, they are pushing full speed to move humanoid robots from tech showrooms to industrial battlegrounds.

At the same time, China has also taken more actions. On March 26, the China Academy of Information and Communications Technology (CAICT), together with more than 40 organizations, released the first industry standard in the field of embodied intelligence. Combined with continuously strengthened policies, faster enterprise implementation, and soaring investor enthusiasm, China is achieving a leap from “following” to “co-running,” and has even begun to challenge for leadership in multiple domains.

Can this revolution—one that overturns future business rules and the industrial ecosystem—allow China to take the starring role?

Global Breakneck Speed: Silicon Valley Giants Move Toward Mass Production and Reshape Future Productivity

No one is treating humanoid robots as a sci-fi concept anymore.

Recently, the synchronized moves by the three Silicon Valley giants have made the footsteps of the mass-production era clearly audible. Every step in their layout points directly to a reconfiguration of future productivity, while follow-through by global capital and local companies has kept the heat on this track rising.

Google is the first to build a “smart brain” for robots, launching two new AI models: Gemini Robotics and Gemini Robotics-ER. The former enables robots to understand new situations without needing dedicated training, while the latter can “understand a complex and dynamic world,” empowering real-world deployment scenarios from a technical standpoint.

Amazon focuses on scenario-based deployment. Within a week, it acquired two humanoid robotics startups—Fauna Robotics—and a logistics robotics company—Rivr. Its strategy is not only to optimize package delivery, but to build “robot service capillaries” that run from factory production lines to the living room, creating a next-generation labor system.

Tesla’s mass-production moves have drawn the most attention. On March 25, its Optimus robot released a related talent recruitment announcement, stating that it would change the global labor and manufacturing industry economic landscape. The goal is to achieve large-scale mass production as soon as possible, and this summer it will open the first humanoid robot production line in human history with an annual output of one million units, pushing mass-production rollout into a more substantive phase.

The Silicon Valley playbook goes far beyond this. On the same day, a humanoid robot Figure03 developed by Figure AI moved into the White House, becoming the first White House humanoid robot made in the United States. It can communicate in multiple languages and autonomously complete household chores, among other capabilities. The company had already raised more than $1 billion in funding half a year earlier, and major players such as Nvidia and LG have publicly backed it, fully showing global capital’s enthusiasm for the humanoid robotics track.

Yuan Shuai, deputy director of the investment department at the China Academy of Urban Development, said that the mass-production moves by Silicon Valley giants and the release of China’s industry standards for embodied intelligence jointly mark that the humanoid robotics industry is moving from the deep waters of technical R&D into a golden window for commercial deployment. Breakthroughs in core technologies support scaled manufacturing, while industry standards set technical requirements and reduce disorderly competition.

However, Gao Heng, an expert from the China Society of Science and Technology News, offered a cautious assessment. He believes that the industry is currently in the prelude to commercialization and a period of partial realization, not the golden window for a comprehensive commercial explosion. The core shift in the industry right now is that various parties are starting to jointly validate whether robots can keep working continuously in real scenarios and whether costs are controllable—not simply whether technical R&D has achieved breakthroughs.

China’s Breakthrough: Multiple Advantages Gain a Foothold, and Key Weaknesses Urgently Need to Be Filled In

When the Silicon Valley giants triggered a mass-production wave, China was not passively following; it had already laid groundwork in advance. With multiple advantages—standards, scenarios, markets, and capital—China has secured its position in the global embodied intelligence track. Still, compared with Silicon Valley giants, China has a gap in core technologies and capabilities, which has become a constraint on further industry development.

In terms of advantages, China’s layout shows distinct local characteristics and the benefits of being early. First, it holds the power to shape the narrative on standards. On March 26, CAICT and more than 40 organizations released the first industry standard in the field of embodied intelligence, building a unified benchmark testing framework. In the early stage of industrial development, this allowed China to seize the initiative in standard-setting.

Second, scenario-based deployment is leading. China’s embodied intelligence development has never stayed at the demonstration stage. It has truly achieved real-world applications. For example, Unitree’s quadruped robot has been deployed in multiple industrial inspection projects, including Zhejiang power substations, Hangzhou underground utility corridors, and the Guangdong Petrochemical base.

At the same time, China has a huge market scale and an active capital environment. In 2025, there were more than 140 embodied intelligence end-to-end equipment companies in China, released over 330 humanoid robot products, and shipped approximately 17,000 units. The market sizes of embodied intelligence and humanoid robots are 5.295 billion yuan and 8.239 billion yuan, respectively.

On the capital side, Unitree Technology’s IPO filing was accepted, becoming the first humanoid robotics stock on the A-share market. Since the beginning of this year, embodied intelligence has seen a surge in large-scale financing, accelerating the process of capitalization. In addition, from January to September 2025, Unitree Technology’s sales revenue from quadruped robots and humanoid robots increased year over year by 182.22% and 6.42 times, respectively—intuitively confirming market potential.

Despite strong momentum, China’s shortcomings in the global competition are also evident.

Multiple experts point out that the core differences between humanoid robots in China and abroad are not in hardware manufacturing, but in data accumulation, model generalization ability, and underlying technical foundations. On the surface, this is reflected in robots’ lack of motion flexibility and insufficient generalization ability.

Yuan Shuai believes that the gap between Chinese and international humanoid robots, at the surface level, is the difference in motion flexibility and generalization ability, and at the root lies in underlying technology, data accumulation, and R&D philosophy. For instance, Google’s RoboCat can achieve flexible generalization locomotion. It relies on long-term technical accumulation—especially sustained investment in big-model algorithms, sensor fusion, and robot dynamics control—so that, backed by massive multi-scenario training data, robots have the ability to learn autonomously and adapt to their environment.

He said that domestic products currently mostly remain at stages of preset actions and fixed-scenario replication. The core weaknesses are: first, a lack of high-quality, large-scale real-scenario training data, leading to insufficient algorithm generalization ability; second, key components such as high-precision servo motors and force sensors depend on imports, constraining motion precision and perception levels.

Gao Heng added that the real gap is whether data, models, system engineering, and closed-loop scenario capabilities can form a coordinated linkage. The goal of top overseas companies is to build intelligent robots that can understand the environment and complete tasks autonomously; the key is to develop robots as sustainable, iterating data products. Generalization ability is inherently a composite capability. It is not that domestic technology lags on a single point; rather, data and scenarios have not formed an iteration flywheel. As a result, robots can only tune parameters on single tasks, making it difficult to become smarter the more you use them.

Renowned finance writer Gao Chengyuan, head of the Jingyuan Influence Research Institute, said that the core gap is concentrated in data accumulation and model generalization ability. Overseas advantages are obvious in transfer learning from simulation to reality and in multi-task universal strategies, supported by long-term investment that builds cross-scenario data closed loops and the capability to develop base models. Domestic efforts still center on preset actions. Fundamentally, this is due to a shortage of high-quality embodied data, and there is a generation-level gap in the compute power and algorithm engineering capabilities required by end-to-end big models.

Unitree Technology also openly acknowledged that key technologies facing the breakthrough needed for large-scale commercial applications in industrial and home scenarios mainly include two difficult issues: embodied big-model capabilities at the “brain” level, and the fine, durable precision of the “dexterous hands.” Among them, the most significant technical challenge is that embodied big models worldwide are still in an early development stage, so generalization ability is insufficient.

The Road to a Breakthrough: Improve Capabilities Through Multiple Dimensions, Balancing Today and the Long Term

Against the backdrop of insufficient data and scenario accumulation, how to enhance robots’ motion flexibility and generalization ability has become the core question for domestic companies to catch up.

Multiple experts, based on the current state of the industry, proposed development paths that combine practicality with foresight. At the same time, they emphasized that enterprises must balance near-term deployment with long-term R&D. Preset actions serve as the entry ticket, while generalization ability is the core competitive barrier.

Wang Peng, a researcher at the Beijing Academy of Social Sciences, suggested two paths for domestic companies to catch up: “scenario anchoring + technical reuse.” On one hand, focus on closed-loop data for vertical scenarios. First lock in standardized scenarios such as industrial welding and material handling. Then use small-scale deployments to obtain proprietary datasets and train embodied models in vertical domains. On the other hand, rely on open-source ecosystem collaboration: leverage the industry standards released by CAICT to promote cross-enterprise data sharing, and jointly train common models based on operational data in a unified format.

Yuan Shuai advised pursuing multiple paths in parallel. On one side, work with universities and research institutions to generate virtual data via simulation and digital twins for training, then transfer to real-world settings. On the other side, open up interfaces to coordinate with scenario partners to run pilot projects, collect real data, and iterate algorithms. Meanwhile, promote anonymous training-data sharing between companies to break data silos, increase investment in the self-development of core components, and support flexible robot motion through hardware breakthroughs.

Gao Heng provided four practical paths: First, obtain data from real scenarios and deeply bind robots to scenarios like factories and warehouses, so robots integrate into real workflows to accumulate data. Second, simulation first with a real-machine closed loop: train strategies in simulation environments first, then fine-tune them in real-world settings to reduce training costs. Third, do task generalization first, focusing on single-task types such as sorting and transportation, achieving generalization to realize commercial value first. Fourth, establish an industry-shared data and standards system to address the problem that interfaces and evaluation systems are not unified, forming an industry-level iteration.

Experts unanimously believe that preset actions and generalization ability are equally important for enterprise development.

Wang Peng said that in the short term, robots with preset actions can already cover the needs of most industrial scenarios, and their costs are only that of generalization-capable robots. But in the long run, generalization ability is the core barrier that determines whether a company can cross industry cycles. As non-standardized scenarios expand—such as home services and emergency rescue—robots that can autonomously adapt to the environment will gradually become mainstream.

Gao Heng agreed as well: preset actions are today’s entry ticket, while generalization ability is tomorrow’s championship ticket. For companies, they cannot give up long-term investment in generalization ability just because they can make money from preset actions today; but they also cannot, conversely, ignore the scenarios that can be deployed right now because they pursue generalization. Get orders first, then train intelligence—that is the more realistic route.

With the market size of China’s embodied intelligence already accounting for about half of the global total, and with deployments and applications realized in industrial and emergency scenarios, going forward, which type of scenario will become the first breakout point for China’s embodied intelligence robots to achieve large-scale commercial adoption?

Gao Chengyuan said that industrial manufacturing will be the first breakout point for China to achieve large-scale commercialization, especially in scenarios such as automotive manufacturing, 3C electronics assembly, and warehousing and logistics. To uncover scenario demand, it is necessary to go deep into the front line of the industry and co-build joint laboratories with leading manufacturing companies. Start by replacing single processes, then gradually expand to end-to-end automation. The key to integrating technology and scenarios is to establish a reverse-driven mechanism of “scenario-definition technology,” so that real production-line needs pull hardware iteration and algorithm optimization—rather than technology leading first and then searching for scenarios.

From “co-running” to “leading globally,” China still needs to break through core bottlenecks in policy, technology, and the industrial ecosystem.

Yuan Shuai suggested that at the policy level, support and funding should be strengthened and intellectual property protection improved. On the technology front, focus on breaking through big-model algorithms and core components, and improve robots’ autonomous learning and generalization capabilities. In terms of the industrial ecosystem, strengthen coordination between upstream and downstream players, accelerate the localization of components, deepen the integration of industry-university-research-use, and promote the commercialization of results. At the same time, actively carry out international cooperation, participate in global standard setting to enhance industry discourse power, and ultimately build a complete embodied intelligence industrial ecosystem to achieve the goal of global leadership.