In October, the China Academy of Information and Communications Technology (CAICT), together with the Department of Electronic Engineering at Tsinghua University and the Embodied Intelligence Testing Laboratory, released the annual Embodied Intelligence Development Report (2026) for the third time. This 82-page report compresses the adjective for the embodied intelligence industry in 2026 into four words: hot yet cold.
On the hot side, domestic financing in the first half of 2026 reached CNY 93.5 billion, completing in half a year what was done in the whole of 2025; global humanoid robot shipments reached 22,000 units, a year-on-year increase of nearly 300%. On the cold side, the report presents a set of sobering data: educational and research institutions account for more than half of the publicly disclosed winning bids, and the average monthly operating days for many robots are less than 10 days; daily rental for performances dropped from CNY 20,000 to the thousands of CNY, and the residual value of second-hand engineering prototypes shrank by nearly 90%.
The report's judgment is straightforward: the market has already priced in the technological value in advance, while the industry is still in the early stage of transitioning from technological validation to commercial validation. This article breaks down the key data and judgments in the report: where the money comes from, where the coldness lies, how far the technology has advanced, who is making money, and a ruler that can measure the true level.
I. The Hot Side: Money, Shipments, and New Products Rising Together
Let's look at capital first. According to statistics from IT Juzi, in the first half of 2026, there were 322 financing events in the domestic embodied intelligence track, totaling approximately CNY 93.5 billion. In just half a year, it exceeded the CNY 73.5 billion for the whole of 2025, an increase of about 5 times compared to the same period last year, with the number of financing events increasing by 137% year-on-year.
Money is concentrating at the top: the top 20 companies by financing amount totaled approximately CNY 55 billion, accounting for 59% of the total financing; among them, humanoid robot body companies accounted for about CNY 31.5 billion, embodied "brain" companies about CNY 14.5 billion, and core components about CNY 5.3 billion. During the same period, more than 800 investment institutions participated, and the number of companies reaching a valuation of CNY 10 billion increased to 20, most of which focused on the embodied brain.
Large overseas financing is concentrated in models and world models: Skild AI and Germany's NEURA Robotics each completed a USD 1.4 billion Series C round, while World Labs and AMI Labs each secured USD 1 billion-level financing for world models. World models became a target for large-scale capital for the first time in 2026.
The product side is also scaling up: 123 new products were released at the World Robot Conference in 2025, increasing to 311 in 2026, more than doubling; in the first half of 2026, global humanoid robot shipments exceeded 22,000 units, a year-on-year growth of nearly 300%, of which five domestic complete machine manufacturers accounted for a combined share of over 86%.
On the policy front, embodied intelligence has been written into the Government Work Report for two consecutive years. The "15th Five-Year Plan" sets the tone, and the real-scene practical training special action provides hard targets: promoting the shift of robots from "demonstration and performance" to normalized "operation mode" by the end of the year, and deploying over 10,000 units. The certainty expectation of the industry has significantly increased.
II. The Cold Side: Four "Not Equals"
In the "Domestic Status" section, the report presents four inequalities: progress in demonstration-level capabilities does not equal real production line operation capabilities; shipment growth does not equal sustained and effective operations; an increase in intended orders does not equal payment collection, repurchases, and a closed commercial loop; and the cost of a single machine does not equal the deployment cost. Behind each set of inequalities lies data.
Technology is cold. Most domestic companies are still reproducing and fine-tuning earlier open models such as π0, π0.5, and OpenVLA, lagging about two generations behind cutting-edge closed-source models. Data utility is also limited: the Dyna-2 model increased pre-training data from 100,000 hours to 1 million hours, but the average success rate for 14 tasks only increased from 45% to 53%—while the training data for most domestic models is still within 100,000 hours, a difference of an order of magnitude.
The true level of cutting-edge models can also be quantified: the Stanford 2026 AI Index shows that the task success rate of robots in real home environments is only 12%, while in simulated environments it is 89.4%; GPT-6 Astra performs sub-millimeter puzzle insertion with a 10% success rate. Understanding it does not equal being able to do it.
Applications are cold. In the first half of 2026, among publicly disclosed embodied intelligence-related winning bid projects in China, educational and research institutions accounted for about 55.5%, government and public platforms about 20.6%, and industrial and technical enterprises only about 21.1%. A considerable proportion of equipment is only powered on for exhibition hall receptions or event demonstrations, remaining largely idle outside of performance schedules; it is not uncommon for the average monthly operating days to be less than 10 days.
The gold content of orders is also being discounted: current public orders are mixed with framework agreements, strategic cooperation, and demonstration purchases. Framework agreements often only stipulate cooperation intentions without purchase quantities, and demonstration purchases carry task attributes rather than market behaviors. When calculating the total account, the price of the complete machine only accounts for 40%—60% of the total cost of ownership (TCO), and the annual comprehensive operation and maintenance cost of a single wheeled-arm or bipedal humanoid robot is about CNY 500,000 to CNY 700,000, far higher than the labor cost for most positions.
Enterprises are cold. In the performance rental market, the daily rental for humanoid robots dropped from about CNY 20,000 in early 2025 to the thousands of CNY in 2026, a drop of 80% to 90%; in the second-hand market, engineering prototypes purchased for CNY 300,000 to CNY 800,000 in 2024—2025 are now only worth CNY 30,000 to CNY 60,000, with the residual value shrinking by nearly 90%.
The books of listed companies tell the same story: Yuejiang Technology's revenue in the first half of 2026 was CNY 316 million, a year-on-year increase of 106.6%, but the net loss expanded to CNY 108 million, with revenue still mainly coming from the traditional collaborative robot business; over 500 listed companies on the A-share market mentioned embodied intelligence in their semi-annual reports, but only a single-digit number disclosed related business revenue. Industry reshuffling has begun—Yixing Robot dissolved just 5 months after its establishment, and Gangtiexia Technology has not received new financing for 32 months.
The report does not shy away from its attitude towards the bubble: the stage from 0 to 1 in the hard technology industry is always accompanied by a bubble. The Internet bubble around 2000 precipitated broadband and data centers, and the capacity expansion of new energy vehicles from 2015 to 2018 precipitated a complete industrial chain. Embodied intelligence is currently in the period of inflated expectations, and will complete the first round of clearing through market competition in the next two to three years.
III. Technological System: Advancing on Four Fronts Simultaneously
In 2025, the industry was still debating technological routes: how to choose between hierarchical architecture and end-to-end, and who would dominate among VLM, VLA, and world models. The conclusion for 2026 is: the dominant architecture has not yet emerged, and the routes are far from converging. The industry has shifted to simultaneously exerting efforts along the four elements of data, models, bodies, and computing power, seeking breakthroughs through system synergy.
On the model front, three lines are running in parallel. First, exploring the scaling law: GeneralistAI's GEN-1, after pre-training with 500,000 hours of data, requires only 1 hour of robot data to learn new tasks such as folding shirts, with a 99% success rate; GEN-1.5 goes further, able to perform 10 new tasks after watching a 3—12 second motion demonstration. Second, multi-route iteration of VLA, world models, and whole-body control: Figure's Helix uses a low-level controller trained with 1,000 hours of data, directly replacing about 110,000 lines of hand-written C++ code. Third, the agent paradigm is landing first: large models are responsible for task decomposition and command, while specialized policies are responsible for execution. NVIDIA's ENPIRE framework enables coding agents to autonomously train dexterous manipulation policies with a 99% success rate.
On the data front, the gap is the biggest bottleneck. Large language models can get trillions of tokens for free from the Internet, but embodied intelligence requires multi-modal interactive data such as vision, motion, and force/tactile feedback, which has never been systematically recorded and can only be produced item by item. Calculations show that for robots to usher in the "ChatGPT moment," tens of millions of hours of effective data are needed. The gap is a business: in the past year, there have been 97 financing events in the embodied data track, involving 34 companies, with disclosed amounts exceeding CNY 20 billion.
On the body front, components are mature, but complete machines remain difficult. Integrated joint efficiency has reached 90%, with a peak torque of 600 Nm; dexterous hands have achieved 22 active degrees of freedom; battery specific energy is 380 Wh/kg, completing hot-swappable battery changes in just over 10 seconds. Basic indicators are getting closer and closer, but the report's judgment is: it is becoming easier to build bodies with similar performance, but it remains very difficult to build products that can adapt to real scenarios and operate stably over the long term.
On the computing power front, cloud-edge-device collaboration has become the standard. The device side is a hard bone to crack: through multi-layer optimization, the end-to-end latency of the π0.5 model on Jetson Thor has been compressed from 278 milliseconds to 29 milliseconds, enabling robots to work in real-time on power-constrained bodies.
IV. Applications: The Three-Step Staircase from "Performing" to "Working"
The first step, cultural, entertainment, and sports events. This is an arena for capability demonstration and a testing ground for technology. The on-site attendance at the World Robot Conference was 271,000 in 2025, reaching 557,000 in 2026; the number of teams participating in the Yizhuang Humanoid Robot Half-Marathon increased from 20 to 102, and the champion's time was shortened from 2 hours and 40 minutes to 50 minutes and 26 seconds, less than a third of last year's time; the 2nd Humanoid Robot Sports Meeting had 666 teams and 2,056 robots participating, with the 100-meter time dropping from 21.50 seconds to 9.32 seconds.
The second step, personal consumption. Toys and companion products have taken the lead in achieving commercial monetization: the BabyAlpha series of robot dogs has accumulated sales of over 25,000 units, with household users accounting for 90%; Casio's Moflin has sold 20,000 units, and Panasonic's NICOBO has exceeded 10,000 units. Beijing has included companion robots and robot dogs in personal consumption subsidies, subsidizing 15% of the selling price, up to a maximum of CNY 1,500—consumer-grade embodied intelligence has officially entered the local policy system for promoting consumption.
The third step, production applications. Industrial manufacturing and logistics warehousing are taking the lead in advancing towards continuous operations: X Square Robot livestreamed sorting 1,816 packages in one hour during a real logistics sorting task; energy and power are beginning to enter substations for inspection, with robots from China Southern Power Grid already piloting in multiple substations in Shenzhen and Guangzhou; medical and elderly care are entering Anzhen Hospital and elderly care stations for trials; in commercial services, Yinhe Tongyong undertakes the entire process of inventory counting, restocking, and picking/delivery in over a dozen pharmacies in Beijing, processing 200—300 orders per store per day.
However, the report also reminds: at this stage, the overall situation is still in the single-station or small-batch pilot stage, with pilots concentrated on low-complexity tasks and closed scenarios. The success rate, stability, and generalization capabilities in open environments cannot yet meet the requirements for scaled production.
V. Industrial Ecosystem: 405 Companies, "Selling Shovels" Making Money First
According to statistics from the Embodied Intelligence Industry Map of the China Artificial Intelligence Industry Development Alliance (as of October 8, 2026), there are 405 domestic embodied intelligence-related companies: 180 in the upstream, 257 in the midstream, and 212 in the downstream, of which 191 companies cover two or more segments simultaneously, accounting for 47.2%—software and hardware synergy, complete machine and component synergy, and product and scenario synergy are the common strategies.
In terms of business models, the differentiation across segments is obvious, and the report's summary is very precise: in most segments, "growth is established, but profitability is not yet established".
| Segment | Monetization Status |
| Data | Closed the loop first. Luming Robot's data supermarket sells 100 pieces of real-scenario data for over CNY 100, and teleoperated real-machine data for hundreds to over CNY 1,000; Guanglun Intelligent saw revenue grow 10 times in 2025 |
| Models | Indirect monetization. Complete machine companies include self-developed models in the pricing of complete machines; model companies charge by subscription or project, and Skild Brain already has overseas manufacturers paying for integration |
| Complete Machines | Dual track of sales and leasing. The gross profit margin of quadruped robots can reach 57.5%; 1X's NEO offers a monthly USD 499 home entry plan; desktop robots at USD 399, with orders exceeding USD 1 million within 6 hours of launch |
| Scenario Services | Initial stage of solution landing. Yinhe Tongyong builds its own pharmacy scenario for replication; Spirit AI signs contracts with JD MALL to take up posts; after-market forms such as robot 4S, 6S, and 9S stores are emerging |
Table: Current Status of Business Models in Various Segments of the Industrial Chain (Compiled based on Chapter 4 of the report)
Open source is also becoming the main line of the ecosystem: as of September 2026, there are over 100 open-source embodied intelligence datasets globally, and the number of open-source releases in 2026 has reached twice the total for the whole of 2025; Physical Intelligence open-sourced the OpenPi repository, and Robbyant open-sourced LingBot-VLA and completed adaptation with mainstream bodies such as Xinghaitu, Songling, and Leju. Trading open source for ecological niche is the new strategy in 2026.
VI. A Ruler: Maximum Real-Machine Success Rate of 24.4%
Based on about 800 publicly published embodied intelligence papers in 2026, the report reviews existing evaluation benchmarks, categorizing them into four types: cognition, navigation, manipulation, and safety. Several numbers are worth remembering:
Simulation test papers are near perfect, but adding disturbances reveals their true form. On the LIBERO benchmark, the self-reported success rates in papers are generally between 97%—99%; however, after LIBERO-Plus adds seven-dimensional disturbances such as camera pose, lighting, and initial position, models that were originally above 95% generally drop to below 30%, with π0 dropping the worst to 6.0%.
From simulation to real machine, the gap is significant. Research from Tongji University found that when representative navigation models are transferred from simulation to real environments, the success rate drops from 61% to 22%; as of September 2026, the highest success rate on the RoboChallenge real-machine evaluation leaderboard is only 24.4%, while human experts achieve 100%.
Single-step reliability does not equal long-horizon reliability. Even if single-step reliability reaches 99%, the overall success rate after continuously executing 100 steps is only about 37%—this is the mathematics of error accumulation, and also the hardest gap between "demonstration" and "delivery".
Standard construction is filling the gap: in December 2025, the "Embodied Intelligence System Framework and Capability Requirements" led by our country became the world's first international standard for embodied intelligence; in February 2026, the ITU established the Embodied Intelligence Standard Focus Group, which has attracted participation from 39 countries and over 120 institutions; two domestic industry standards have achieved "zero breakthroughs" in the direction of evaluation and dataset quality. The experience of the communication industry is that the years before commercialization are precisely the golden period for standard layout, and embodied intelligence is currently in this window.
VII. Conclusion: Like Electric Vehicles in 2015, Not the Metaverse in 2020
The most valuable part of the report is that it does not take sides to "hype" or "bash," but lays out both the hot and cold sides: the bubble truly exists—rents dropping by 90%, prototype residual values dropping by 90%, revenue increasing but profits not, these are all scenes of bubble squeezing; but technological progress is equally real—marathon times shrinking to a third, 100 meters run in under 10 seconds, device-side latency compressed to 29 milliseconds, real-machine policy success rate reaching 99%, these are also facts.
The path of industrial evolution also provides a frame of reference: technologically, it may fluctuate periodically like AI large models; the industrialization process can be analogous to new energy vehicles; the commercialization path is closer to autonomous driving—entering operations first when partial capabilities are available, and then gradually expanding the scope.
Application expansion will follow the "80/20 rule": 80% of resources are concentrated in deterministic scenarios such as inspection, handling, sorting, and quality inspection, serving as the "beachhead" for priority breakthroughs; 20% of exploration is left for open environments such as home services and elderly care. Robots do not need to wait for the emergence of a "perfect general-purpose humanoid" to enter the industry, maturing while deploying is a more realistic path.
For practitioners, the report's inspiration is simple but useful: do not bet on a single segment; the focus of competition has shifted from model performance or hardware parameters to the full-stack synergy of "data—model—body—scenario"; in the next two to three years, the industry will complete the first round of clearing, and only companies that can solve the four problems of generalization, reliability, cost, and data closed-loop will remain at the table.
The content and data in this article are all cited from the "Embodied Intelligence Development Report (2026)" (released in October 2026) by the Institute of Artificial Intelligence of the China Academy of Information and Communications Technology, the Department of Electronic Engineering at Tsinghua University, and the Embodied Intelligence Testing Laboratory.