Embodied Intelligence: China’s Next Frontier in AI

Artificial intelligence is undergoing a fundamental transformation in China, moving from the digital realm into the physical world. Known as embodied intelligence (具身智能), this emerging field combines AI with physical form, enabling machines to perceive, learn from, and act upon real-world environments in a continuous feedback loop of perception, cognition, decision, action, and adaptation.

President Xi Jinping has framed this shift as a strategic priority, noting that artificial intelligence is “a new engine for world economic growth and an accelerator for the transformation of old and new growth drivers,” moving from the “digital world” to the “physical world,” as reported by Xinhua News.

A Policy Priority and a Trillion-Yuan Market

China’s policy apparatus has moved decisively to elevate embodied intelligence. The country’s “15th Five-Year Plan” explicitly lists it as one of six future industries to be developed, alongside quantum technology, biological manufacturing, brain-computer interfaces, and 6G communications. The technology was first written into the Government Work Report in 2025, and the 2026 report went further, proposing to “build a new form of intelligent economy” with embodied intelligence named as a key future industry.

The market potential is staggering. According to the China Development Report 2025, compiled by the State Council Development Research Center, China’s embodied intelligence industry is projected to reach 400 billion yuan (approximately $56 billion) by 2030 and exceed 1 trillion yuan by 2035, as Jiemian News reported.

Investment is already flowing. In March 2026, Galaxy General Robotics completed a 2.5 billion yuan funding round with investors including the National AI Industry Investment Fund and Sinopec - marking the first time a national-level major fund has publicly entered the embodied intelligence field, according to Guangzhou Daily.

What Exactly Is Embodied Intelligence?

Embodied intelligence is not simply about humanoid robots, though humanoid form is one possible carrier. The key distinction lies in whether an intelligent system can truly enter real environments and form a closed “perception-cognition-decision-action-feedback” loop, continuously adjusting its behavior and improving its internal models through learning.

Zheng Nanning, a Chinese Academy of Engineering academician and director of the Institute of Artificial Intelligence and Robotics at Xi’an Jiaotong University, explains the concept through the analogy of infant development. “The essence of embodied intelligence is not to make machines ‘like humans,’ but to let intelligence truly enter the real world, forming continuous learning and adaptation capabilities in complex, uncertain environments,” Zheng said in an interview published by People’s Daily.

An embodied intelligence system typically consists of three components: a perception system (vision, touch, force sensing) that acquires environmental information, a decision system that forms understanding and planning, and an execution system (robotic arms, mobile platforms) that acts on the physical world. These three components form a closed loop, enabling the system to learn through action rather than relying solely on offline data training.

From ‘Understanding the World’ to ‘Growing in the World’

Zheng contrasts embodied intelligence with current large language models, which he describes as “intelligent agents in a study” - skilled at language generation and knowledge expression but unable to directly act on the physical world. “Embodied intelligence is not a simple extension of existing models, but pushes us to re-examine the role of body, action, and environment in the formation of intelligence,” he said.

The academician outlines four key capabilities required for embodied intelligence to truly “enter the world and act within it”:

• World model formation: Building internal simulation environments that allow prediction and reasoning before action, shifting from passive response to proactive decision-making.
• Multimodal representation: Integrating visual, language, tactile, and auditory information into a unified semantic space for holistic understanding.
• Causal reasoning: Establishing stable “action-result” causal chains to avoid AI hallucinations - a critical gap in current large models that rely primarily on statistical correlations.
• Integrated generation and action: Developing end-to-end systems that unify vision, language, and action modeling, enabling robots to generate action strategies from natural language instructions.

Historical Roots and Rapid Acceleration

Embodied intelligence is not a new concept. Alan Turing first proposed the idea in 1950, envisioning machines with bodies and perception abilities. China’s 863 Program, launched in 1986, explicitly proposed developing precision assembly robots, underwater operation robots, and robots for complex hazardous environments - early prototypes of embodied intelligence that had to operate in real-world conditions.

But the field has accelerated dramatically in recent years, driven by advances in large model technology, improved hardware, and major tech company investment. At the China Embodied Intelligence Conference (CEAI 2025) in March 2025, the “Fifteen Key Directions for Embodied Intelligence” and the China Association for Artificial Intelligence Embodied Intelligence White Paper were released - the first systematic roadmap for embodied intelligence technology development in China, as Dazhong News reported.

Tan Tieniu, a Chinese Academy of Sciences academician, told the conference that “embodied intelligence is not only a form of artificial intelligence, but also the only way for the sustainable development of AI. Stacking computing power is not sustainable; embodied intelligence based on biological inspiration has broad prospects.”

Standards, Commercialization, and Real-World Applications

China has moved quickly to establish a regulatory and standards framework. On February 28, 2026, the first national “Humanoid Robot and Embodied Intelligence Standards System (2026 Edition)” was officially released at the Humanoid Robot and Embodied Intelligence Standardization (HEIS) Annual Conference in Beijing. The system covers six aspects: basic commonality, brain-like intelligence and computing, limbs and components, complete machines and systems, applications, and safety ethics, as CCTV News reported.

Liang Liang, secretary-general of the MIIT Humanoid Robot and Embodied Intelligence Standardization Technical Committee, emphasized the importance of scenario-based standards: “By setting application standards for different scenarios, we can help these technologies move from pilot laboratory products to commercial deployment. This can also improve product quality and market trust, helping to build an open, innovative, and sustainable industrial ecosystem.”

The applications span multiple sectors. In industrial manufacturing, intelligent systems are evolving from execution tools to intent-driven systems that can monitor product quality in real time, participate in production scheduling, and enable full-process visualization through digital twin technology. In healthcare and rehabilitation, AI systems are becoming essential assistants for minimally invasive surgery and recovery training. In urban governance and public safety, drones and ground robots are performing infrastructure inspections in high-risk environments. And in transportation, autonomous driving systems represent one of the closest forms of embodied intelligence to large-scale deployment.

TrendForce predicts that 2026 will be a key year for humanoid robot commercialization, with global shipments expected to grow more than sevenfold year-on-year, exceeding 50,000 units.

Challenges Ahead

Despite the momentum, significant challenges remain. Zheng identifies several critical obstacles: the black box problem of insufficient interpretability in model decision processes, the contradiction between computing power requirements and edge deployment, an immature industrial ecosystem with evolving standards and interface specifications, and a data bottleneck - the lack of high-quality multimodal data as “fuel.” Additionally, narrowing the gap between simulation training and real-world deployment remains an unsolved problem.

“From data-driven large models to embodied intelligence oriented toward the physical world, artificial intelligence is undergoing a continuous evolution from the symbolic world, data world, to the real world - from ‘understanding the world’ to ‘growing in the world,’” Zheng said.

What to Watch For

As China pushes forward with embodied intelligence, several developments bear watching: the rollout of scenario-based standards and their impact on commercialization timelines, the progress of major funding rounds and their effect on industry consolidation, and whether breakthroughs in multimodal integration and causal reasoning can overcome the current technical limitations.

The next few years will be decisive. With policy support, substantial capital, and a rapidly maturing industrial ecosystem, China’s embodied intelligence sector is positioned to become a defining force in the global AI landscape - one that moves intelligence from the screen into the physical world.