Embodied Intelligence: From "Future Industry" to "New Growth Point", How to Cross the "Valley of Death" of the Trillion-Yuan Track

On August 26, 2026, at the themed press conference series "Kickstarting the 15th Five-Year Plan" hosted by the State Council Information Office of China, the Ministry of Industry and Information Technology clearly stated that it will "promote future industries including quantum technology, biomanufacturing, hydrogen energy and nuclear fusion energy, brain-computer interface, embodied intelligence, and 6th Generation Mobile Communication (6G) to become new economic growth points".

Among the list of six emerging pillar industries and six future industries, embodied intelligence occupies a unique position — it is not only categorized as a "future industry", but also the one closest to the threshold of "new growth point".

This is not an accidental policy preference, but an inevitable reflection of industrial reality. In 2026, the global shipment volume of humanoid robots is expected to reach 62,500 units, marking a year-on-year increase of approximately 247%, with Chinese manufacturers accounting for more than 80% of global shipments; the broad-sense embodied intelligence market scale is projected to exceed 1 trillion yuan in 2026; the total financing of the domestic track in the first half of the year reached 93.5 billion yuan, a 5-fold year-on-year increase.

Figures do not lie — embodied intelligence is standing at the historic critical point of transitioning from "technology verification" to "large-scale commercial application". However, the path from "future" to "growth" is never a smooth one.

Three Battles in the Trillion-Yuan Track

If the embodied intelligence sector in 2025 was still competing on "who can stand up and walk", the competition in 2026 has fully escalated into three parallel ongoing battles.

The first battle: the battle for the "brain". Industry consensus is taking shape: the competition of embodied intelligence is shifting from "competing on hardware" to "competing on models and data". VLA (Vision-Language-Action Model) and the world model have become two main technical routes, and an increasing number of enterprises realize that the two are not mutually exclusive, but have distinct complementary space.

For instance, KOM 3.0, the world's first service industry VLA architecture integrating latent space world model released by Qianglang Intelligence, allows robots to deduce physical results internally before executing actions; Magic-VLA K02, the general embodied large model demonstrated by Magic Atom, constructs a closed technical loop of "scene setting questions - model solving questions - data feedback".

Yet the "brain" remains a scarce asset. Industry data shows that the average success rate of top models on the simulation leaderboard is only 8.80%, and that on the real machine leaderboard is merely 12.8%. This means that even in laboratory environments, the vast majority of tasks still cannot be reliably completed by general models. The Harness VLA released by the team of Professor Yu Chao from Tsinghua University in July attempts to solve the post-failure reset problem of VLA in contact-intensive manipulation by introducing the Harness Layer — which precisely illustrates that current VLA models are still far from being "reliable".

The second battle: the transformation of "limbs". The "brain" is catching up, but the "limbs" are already running. Since 2026, the localization rate of the three core components including reducers, servo systems and controllers has risen to 75% to 90%. The harmonic reducers of GreenHarmonic, as well as the self-developed controllers and servo systems of Estun, have all been supplied in batches to large domestic and foreign enterprises. The self-development and self-production rate of core components of Unitree Robotics has exceeded 90%. Chinese manufacturing is moving from "scale dividend" to "technology dividend".

But the "limbs" are not free from hidden worries. Practical challenges such as the supply-demand mismatch between complete machines and components, poor transformation of technological achievements, and barriers in industry-finance docking remain prominent. The mean time between failures of joint modules still lags behind that of traditional industrial equipment, and temperature rise control remains a key engineering problem. Although the cost of a domestic humanoid robot complete machine has dropped to the 100,000-yuan level, if you examine its core components closely — reducer, frameless torque motor, force sensor — you will find a reality: the torso is in place, but the "brain" is still pending.

The third battle: the battle for "scenarios". If the first two battles determine "whether it can be done", the third battle determines "whether there is market demand". At the 2026 World Robot Conference, a clear signal emerged: the industry is shifting from "demonstrating capabilities" to "verifying value". Compared with the demos of folding clothes and making coffee displayed by various enterprises last year, most enterprises this year have presented real orders or POC results.

Starward Era announced that it has taken the lead in realizing PMF (Product-Market Fit) in the logistics industry, and has achieved normal operation in more than 10 logistics centers across 5 provinces and cities in cooperation with SF Express and China Post; the wheeled humanoid robot of Magic Atom has undertaken real production tasks including material handling and loading/unloading in the Dreame Intelligent Manufacturing Factory; JD plans to purchase 3 million robots in the next 5 years. The special real-scene practical training action jointly carried out by the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission aims to refine more than 100 high-value application scenarios by the end of 2026, and drive the formation of a ten-thousand-unit scale deployment capacity.

The industry is rapidly switching from "R&D state" to "deployment state".

Where is the Bottleneck?

Although the three battles mentioned above are in full swing, the entire industry is facing a common "stuck neck" problem: data.

This is not an ordinary "lack of data". Digital AI has nearly unlimited text, image and video data on the Internet for training, but embodied intelligence requires interactive data from the real physical world — every action of a robot grabbing a cup, screwing a screw, or folding a piece of clothing involves the complex coupling of visual understanding, spatial positioning, contact judgment and motion control. These data are "irreplaceable", and are essentially different from Internet language and visual data.

How large is the gap? Data shows that the current compliant data of real physical interaction scenarios in China is only 500,000 hours, while the commercial deployment of robots requires tens of millions of hours of data, with a gap of over 99%. This is a gap ranging from the order of "hundred thousand" to "ten million".

The cost of data collection is also staggering. The collection cost of 30-second real machine operation data is 10 to 15 yuan; collecting 1 hour of data costs around 1,000 yuan, and gathering 200,000 hours of data requires more than 200 million yuan. If the pre-training of millions of hours of real machine data is completed, the input cost will reach the order of 1 billion yuan. Wang Xingxing, founder of Unitree Robotics, stated bluntly that insufficient generalization ability is currently the biggest bottleneck of embodied intelligence — "the same type of problem can be solved multiple times", but retraining is required when the scenario changes. He calls it "the biggest bottleneck of embodied intelligence across the world" — it is not the technical short board of a single enterprise, but a common dilemma of the whole industry caused by the lack of a unified data base.

Inconsistent data standards also lead to a large amount of raw data cannot be directly used for training, and the whole industry is facing a severe "data famine". More than 70 embodied intelligence training grounds have been built nationwide, but they are facing common challenges including homogenization of scenario tasks, limited actual data production capacity, and insufficient data diversity and circulation. "Collectable data scale" is not equivalent to "trainable data scale".

Breakthrough Path — Three Key Signals

The data dilemma is not unsolvable. In 2026, three key signals are changing the rules of the game.

The first is the special action of real-scene practical training. In June 2026, the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission of the State Council jointly launched the "Special Action for Real-scene Practical Training of Humanoid Robots and Embodied Intelligence". The action focuses on key scenarios in industrial, service and special fields, and deploys 6 key tasks including building real-scene practical training spaces, forming innovative application consortia, and tackling practical operation skills. Its core logic is: let robots "work" in real scenarios, accumulate data, iterate models and verify products in the process of working.

The significance of this action goes far beyond policy endorsement. Through intensively constructed, standardized managed and resource-sharing practical training spaces, repeated scenario construction and resource waste can be effectively avoided, the efficient reuse of scenarios, data, computing power and technological achievements can be promoted, and the trial-and-error cost of enterprises can be greatly reduced. From "usable" to "user-friendly", real-scene practical training is a key leapfrogging measure.

The second signal: exploration of data co-construction and sharing mechanism. The industry is waking up. JD announced that it will collect more than 10 million hours of human real-scene data and more than 1 million hours of robot body data within two years; Securina Intelligent established an embodied intelligence data subsidiary, relying on more than 500,000 hours of real machine data to build a real-world data infrastructure. Xie Shaofeng, Chairman of the OpenAtom Open Source Foundation, called for "establishing a unified data base" to solve the data island problem caused by heterogeneous data formats.

The path proposed by Fan Haoqiang, co-founder of Forcelink Intelligence, is quite representative: the core breakthrough is "real machine data closed loop + embodied native architecture + simulation and synthetic data linkage" — build physical intuition with high-quality real machine data, use physics engines to generate massive extreme scenarios to supplement long-tail data, "start running" in real scenarios first, and continuously absorb feedback for self-iteration.

The third signal: the "15th Five-Year Plan" policy closed loop has initially taken shape. From the special real-scene practical training action in June to the clear statement at the press conference of the State Council Information Office in August, the policy signal is quite clear: embodied intelligence is no longer just a "forward-looking layout", but a "new growth point". Xin Guobin, Vice Minister of the Ministry of Industry and Information Technology, also emphasized at the press conference that it will "tackle cutting-edge technologies including brain-inspired intelligence and world models, and promote the iteration of intelligent terminals such as humanoid robots and brain-computer interfaces". From technological tackling to scenario deployment, from data collection to large-scale implementation, the policy closed loop is taking shape.

If a summary is to be made, it is that the "15th Five-Year Plan" story of embodied intelligence is essentially a story of how an industry crosses the "valley of death".

The so-called "valley of death" refers to the most dangerous journey of new technologies from the laboratory to the market — the technology is not yet mature, the cost has not yet decreased, and the market has not yet accepted it, countless innovative enterprises have failed at this stage. Embodied intelligence is at this juncture: the technical route has not yet converged (VLA or world model?), the cost has not yet dropped to the economic inflection point (calculation shows that the complete machine price needs to be reduced to less than 400,000 yuan and the operation efficiency reaches more than 75% to have overall economy), and the business model has not yet been verified (hardware sales or software subscription?).

But all signs in 2026 show that embodied intelligence is accelerating through this valley of death. Shipments are increasing, costs are decreasing, scenarios are expanding, and policy support is intensifying. The China Business Industry Research Institute predicts that the shipment volume of humanoid robots in China is expected to reach 380,000 units in 2030, with the market scale exceeding 20 billion yuan. The broad-sense embodied intelligence market has entered the trillion-yuan track.

The breakthrough of embodied intelligence is not just the addition of a new emerging industry. It is the only industry that spans three technical systems: "brain" (AI model), "cerebellum" (motion control) and "limbs" (precision components). Its breakthrough will directly drive the coordinated upgrading of multiple pillar industries such as integrated circuits and intelligent robots. As Xu Xiaolan, Chairman of the Chinese Institute of Electronics, pointed out, embodied intelligence has a multiplier effect of 1:10.

When a robot moves from the exhibition stand into factories, warehouses and homes, what it changes is not only the valuation logic of an industry, but also the underlying path of "Made in China" transforming to "Intelligent Manufacturing in China". From "future industry" to "new growth point", embodied intelligence is answering the most fundamental question with actions: what will define China's next-generation manufacturing industry?

This article is from "Sice Strategy Think Tank", and is authorized to be released by 36Kr.