China’s Humanoid-Robot Boom: From Factory Demonstrations to Industrial Deployment

China has become the main arena for humanoid-robot development. More than 140 Chinese companies were reported to be working on humanoid machines by 2025, with more than 330 models released, according to figures cited from the Ministry of Industry and Information Technology (MIIT). [4] The scale is significant because the country combines three ingredients that are difficult to match elsewhere: a deep electronics and manufacturing supply chain, large industrial customers, and public funding aimed at artificial intelligence and advanced robotics.

The strategic goal is not simply to build robots that look human. It is to create machines that can use existing factories, tools and workspaces without forcing companies to redesign every production line. In theory, a bipedal robot could move between stations, pick up different objects, open doors, inspect parts and cooperate with people. In practice, the commercial test is much stricter: a machine must repeat a task safely, quickly and cheaply over long periods.

Figure 1. Factory testing at AgiBot, Shanghai. Image credit: Reuters; see the original report and licensing information in [1].

A market moving from spectacle to supply

Available market estimates differ because some count shipments, while others count installations or research prototypes. Counterpoint data reported by the South China Morning Post put global installations at 16,000 units in 2025, with China responsible for more than four out of five. The same report projected more than 100,000 installations by 2027, with logistics, manufacturing and automotive applications expected to represent 72% of the total. [3]

IDC reported a broader estimate of more than 18,000 global shipments in 2025 and projected over 510,000 units by 2030. It also expected Chinese vendors to account for about 95% of 2025 shipments. [2] These numbers should not be read as proof that hundreds of thousands of productive workers are already operating in factories. IDC noted that more than 85% of 2025 deployments were still concentrated in performances, education, data collection and guided tours, while industrial pilots were beginning to expand. [2]

Indicator	Reported figure	What it suggests
Global humanoid installations in 2025	16,000	Commercial and research activity is growing, but the installed base remains small.
China’s share of 2025 installations	More than 80%	Chinese firms currently have strong access to domestic demand and production capacity.
Projected global installations in 2027	More than 100,000	Logistics, automotive and manufacturing may become the principal demand centers.
Projected global shipments in 2030	More than 510,000	A large market is possible if reliability and unit economics improve.
Chinese humanoid manufacturers by 2025	More than 140	Competition is broad, with substantial risk of consolidation.

The companies leading this push include AgiBot, Unitree Robotics, UBTECH, Leju Robotics and several newer firms linked to electric vehicles, consumer electronics and industrial automation. AgiBot has emphasized a broad product range, open-source elements and data collection. Unitree has built its reputation around dynamic movement, compact designs and balance control. The difference matters: one route prioritizes task coverage and ecosystem scale, while the other places visible emphasis on agile locomotion.

Why China can build at scale

China’s advantage begins with components. Reuters reported that analysts and startups estimated Chinese manufacturers could produce up to 90% of the components needed for humanoid robots. [1] Motors, reducers, batteries, sensors, controllers, structural parts and assembly services can often be sourced within a dense regional network. That shortens development cycles and allows manufacturers to test several hardware versions without depending on a distant supplier base.

Cost is the second advantage. Reuters cited a Bank of America estimate that the average bill of materials for a humanoid could be about $35,000 by the end of 2025 and potentially fall to $17,000 by 2030 if most components are sourced in China. [1] A lower component bill does not automatically create a profitable product; software, maintenance, safety certification, training and batteries also carry substantial costs. It does, however, make pilot programs easier to finance.

Production capacity is beginning to reflect this industrial base. Global Times reported that a Leju Robotics line in Guangdong had reached an annual capacity of more than 10,000 units, with one robot leaving the line roughly every 30 minutes. [4] Such figures describe capacity rather than guaranteed demand, but they show that some Chinese suppliers are designing production systems for thousands of machines rather than a few laboratory prototypes.

Public policy reinforces private investment. Reuters reported subsidies, municipal funds and government procurement, including a rise in state procurement of humanoid robots and related technology from 4.7 million yuan in 2023 to 214 million yuan in 2024. [1] This support can reduce early commercial risk, attract component suppliers and create test environments. It can also intensify competition before the most durable business models are clear.

Endurance, balance and speed

Humanoid robots must solve a difficult mechanical problem: they are tall, top-heavy systems with many joints, narrow feet and limited energy reserves. A factory machine that loses balance, overheats or requires a battery change every few minutes is not a useful replacement for a conventional industrial arm.

Recent demonstrations point to progress in three connected areas. First, balance control has improved through better inertial measurement, force sensing, motor control and reinforcement learning. Unitree’s products have received particular attention for dynamic walking and running. Second, energy management is becoming a design priority. IDC identified hot-swappable batteries, lightweight structures, intelligent power distribution and liquid cooling as important enablers of longer operation. [2] Third, locomotion is becoming faster and more adaptable, with machines learning to handle slopes, turns, uneven surfaces and moving obstacles.

A 2026 Beijing humanoid half-marathon placed these capabilities under sustained stress. The event included more than 100 teams, and nearly 40% competed in an autonomous-navigation category, according to Global Times. [4] Long-distance running is not the same as factory work, but it tests battery endurance, gait stability, thermal control and recovery from small disturbances.

Unitree also said its H1 reached approximately 10 meters per second in a sprint test. The figure is a company claim reported by CGTN and Global Times, not an independent benchmark for routine operation. [4] [5] At that speed, a robot needs rapid posture correction and high power density. The more important industrial question is not whether a machine can sprint for a short clip, but whether it can maintain safe, predictable motion through a full shift.

Figure 2. Humanoid robot during a half-marathon event. Image credit: Reuters; reference image from the event coverage in [4].

The missing link: embodied intelligence

Hardware alone cannot make a general-purpose worker. A useful humanoid needs to perceive objects, understand instructions, plan a sequence, control its body and learn from failure. This combination is often called embodied AI because intelligence is evaluated through action in a physical environment rather than through text or images alone.

AgiBot’s Shanghai data-collection facility illustrates the challenge. Reuters reported that about 100 robots, operated by roughly 200 people, repeatedly performed everyday tasks such as folding clothing, opening doors and preparing a sandwich. [1] The purpose was to collect targeted data for training. Physical tasks generate far less readily available data than internet text, so companies must create supervised demonstrations, simulations and real-world trials.

China’s large manufacturing base can help close this data gap. Robots can be tested in warehouses, automotive plants and logistics centers, producing examples of successful grasps, failed movements, collisions, recovery behavior and task completion times. The risk is that data collected in one tightly controlled plant may not transfer to another site with different lighting, tools, floor surfaces or safety rules. Generalization remains one of the central technical hurdles.

IDC reported that 38% of teams in the 2026 half-marathon used fully autonomous navigation, combining vision, LiDAR, satellite positioning and inertial sensors. [2] This is a useful sign of system integration, but navigation in an open course still differs from manipulating delicate parts at a workstation. Industrial deployment will require reliable hand control, force feedback, fault detection and simple interfaces for human supervisors.

Manufacturing first, service applications next

Manufacturing is the strongest near-term use case because the environment can be structured and the economic value of repetitive tasks is measurable. Early pilots have focused on quality inspection, material handling, assembly, picking, palletizing, machine tending and warehouse movement. [1] [2] These tasks are attractive when labor is scarce, ergonomics are difficult or production needs to run for extended hours.

Humanoids will not replace every industrial robot. A fixed robotic arm remains faster and more precise for a stable, repetitive operation. A humanoid becomes more compelling when a plant changes products often, has many small tasks or needs a mobile worker that can use human-designed equipment. The winning architecture may therefore combine fixed automation for speed with humanoids for flexibility.

Service applications are broader but harder to monetize. Retail guidance, food service, reception, education, inspection and public demonstrations can generate visibility and useful interaction data. Yet these settings expose robots to unpredictable speech, children, crowds and unusual objects. A machine that performs well on a stage may still need close supervision in a hotel, hospital or shopping center.

Robot-as-a-Service models could lower the entry barrier. IDC reported that user acceptance of leasing and subscription models doubled year over year. [2] Customers may prefer to pay for task availability and maintenance rather than purchase an expensive machine outright. Suppliers, in turn, would have an incentive to improve uptime, remote diagnostics and software updates.

Risks and the next test of credibility

The main risk is a gap between production capacity and productive demand. A factory can assemble thousands of robots, but customers will adopt them only when the total cost of ownership is competitive and safety is demonstrable. Short videos of acrobatics or racing can show capability, yet they do not reveal maintenance intervals, failure rates, battery degradation, software supervision or performance across thousands of cycles.

Employment is another concern. Chinese officials and state media have presented robotics as a way to address demographic pressure and preserve industrial competitiveness. [1] The transition could still displace some repetitive jobs before new technical and supervisory roles become available. Responsible deployment will require worker retraining, transparent safety procedures and clear rules for incidents involving autonomous machines.

Standards should become increasingly important. MIIT established a technical committee for humanoid robots and embodied intelligence in late 2025, followed by a system framework released in March 2026, according to Global Times. [4] Common testing methods could help buyers compare battery life, task success, balance recovery, noise, collision behavior and cybersecurity instead of relying on promotional demonstrations.

China’s lead is therefore best understood as an industrial position, not a guarantee of final victory. Its companies have scale, suppliers, funding, data and a large domestic market. The decisive phase will come when robots leave carefully staged demonstrations and prove that they can work safely, economically and continuously alongside people. If they do, humanoids may become a flexible layer of automation for factories and selected service settings. If they do not, the current boom will remain an impressive collection of prototypes, pilots and highly visible experiments.

References

1. Reuters, “China’s AI-powered humanoid robots aim to transform manufacturing”, 13 May 2025.

2. IDC, “From Task Execution to Value Creation: What the 2026 Humanoid Robot Half Marathon Reveals About Industry Progress”, 23 April 2026.

3. South China Morning Post, “China dominates global humanoid robot market with over 80% of installations”, 16 January 2026.

4. Global Times, “Humanoid robot half-marathon approaches; China’s robotics sector grows as new product hits 10 m/s sprint speed”, 12 April 2026.

5. CGTN, “Humanoid robot reaches 10 m/s in test, company says”, 13 April 2026.

Video references

1. Unitree H1 speed test. CGTN video: “Humanoid robot reaches 10 m/s in test, company says” — a short report on the company’s claimed sprint speed. 

2. Humanoid half-marathon coverage. Reuters coverage of China’s humanoid-robot half-marathon — reference footage and reporting on endurance, gait and competition conditions.

3. Manufacturing and embodied-AI training. Reuters report on AgiBot’s data and factory program — visual reference for production-line tests and supervised task collection.