China has become one of the most prolific countries in humanoid robot development, with more than a dozen startups and established manufacturers racing to produce bipedal machines that can walk, manipulate objects, and respond to spoken commands. Companies like Unitree Robotics, UBTECH, Fourier Intelligence, and RobotEra have unveiled models ranging from compact research platforms to human-height machines aimed at factory floors, while government policy has explicitly targeted humanoid robots as a strategic technology. The pace of progress, particularly in transferring AI-trained control policies from simulation to physical hardware, has made the Chinese humanoid ecosystem a subject of serious attention in global robotics.
Who Is Building China’s Humanoid Robots
The Chinese humanoid robot landscape includes both well-funded startups and divisions of larger technology firms. Unitree Robotics, based in Hangzhou, is among the most visible. Originally known for quadruped robots that competed with Boston Dynamics’ Spot, Unitree moved into humanoid territory with the H1 and then the more compact G1, a roughly 1.3-meter-tall general-purpose humanoid that has become a popular research platform in labs worldwide. The G1 has appeared in multiple peer-reviewed studies on locomotion control, fall safety, and energy modeling, making it something of a standard testbed for humanoid research.
UBTECH Robotics, headquartered in Shenzhen, is one of the older players, with its Walker series of humanoid robots dating back several years. The company has focused on service-oriented applications and has demonstrated its robots in settings such as reception desks and guided tours. Fourier Intelligence, also from Shanghai, has taken a different approach with its GR series, designing humanoid robots with a rehabilitation and healthcare angle before broadening into general-purpose platforms. Agibot, backed by a prominent Chinese tech executive, has drawn attention for ambitious goals around embodied AI. And Galbot, a newer entrant, has focused on dexterous manipulation and mobile platforms that combine humanoid upper bodies with wheeled bases.
RobotEra, a Beijing-based startup, has developed the XBot line of humanoid robots in two sizes. The XBot-S stands about 1.2 meters tall and the XBot-L reaches 1.65 meters, both designed as platforms for reinforcement-learning research with an emphasis on walking in real-world environments without manual tuning after simulation training.1arXiv. Humanoid-Gym: Reinforcement Learning for Humanoid Robot with Zero-Shot Sim2Real Transfer Meanwhile, several established Chinese robotics firms that traditionally focused on industrial arms or logistics robots have announced humanoid prototypes, adding to the sense that humanoid form factors are becoming a strategic priority across the industry.
Government Strategy Behind the Boom
China’s central government has played a direct role in accelerating humanoid robot development. In late 2023, the Ministry of Industry and Information Technology released guidelines that identified humanoid robots as a disruptive technology and set an aggressive timetable: initial mass production by 2025 and broader commercial deployment by 2027. These targets are part of a larger industrial strategy that views humanoid robots as converging with advances in artificial intelligence, high-end manufacturing, and semiconductor design.
Provincial and municipal governments have followed with their own incentive programs. Beijing, Shanghai, Shenzhen, and several other cities have created dedicated robotics industrial parks, offered subsidies for humanoid robot companies, and established testing zones where robots can operate in semi-public environments. The cumulative effect is a funding environment where humanoid robot startups can access both venture capital and government grants, lowering the barrier to building and testing physical prototypes. This stands in contrast to the United States and Europe, where humanoid development is driven primarily by private capital with less coordinated government industrial policy.
The strategic framing is not subtle. Chinese policy documents treat humanoid robots as both an economic opportunity and a response to demographic pressure, particularly an aging population and a shrinking manufacturing workforce. Whether the aggressive production timelines will hold is another question, but the policy signal has clearly shaped investment decisions and hiring patterns across the Chinese robotics ecosystem.
Teaching Robots to Walk
Walking is the foundational challenge for any humanoid, and Chinese research groups have invested heavily in using reinforcement learning to solve it. The basic approach involves training a virtual robot inside a physics simulator to walk across varied terrain, then transferring that learned behavior onto the physical hardware. This “sim-to-real” pipeline avoids the expense and danger of training on a real robot that could fall and break itself, but it requires the simulation to be accurate enough that the learned behavior still works on the actual machine.
Researchers at RobotEra developed Humanoid-Gym, a reinforcement-learning framework built on Nvidia’s Isaac Gym simulator, specifically designed to make this transfer as seamless as possible. The framework was validated on both the XBot-S and XBot-L robots, achieving what the team described as zero-shot transfer, meaning the control policies worked on the real robot without any additional tuning after simulation training.2arXiv. Humanoid-Gym: Reinforcement Learning for Humanoid Robot with Zero-Shot Sim2Real Transfer That kind of result is significant because it suggests the simulation fidelity and training methods are mature enough to skip the painstaking real-world fine-tuning that used to be required.
A separate line of research has tackled the problem of reward design in reinforcement learning. Training a robot to walk means defining what “good walking” looks like mathematically, and getting those reward signals right has traditionally required laborious hand-tuning by engineers. A study using the Unitree G1 demonstrated that automatic reward learning methods can substantially outperform manually designed reward functions, improving both the efficiency of training and the quality of the resulting walking behavior in real-world tests.3PubMed Central. Deep Reinforcement Learning for Real-World Humanoid Robot Locomotion Control with Automatic Reward Learning By automating what used to be one of the most tedious parts of the pipeline, this kind of work accelerates how quickly new robots can be brought from prototype to walking demonstrator.
Embodied Intelligence and the Role of Large AI Models
Walking is only part of the picture. The broader ambition for Chinese humanoid robots, as in the global field, is “embodied intelligence”: robots that can understand spoken instructions, perceive their environment through cameras and sensors, and then carry out physical tasks without step-by-step programming. This is where the recent explosion in large language models and vision models enters the humanoid story.
The prevailing research direction is toward Vision-Language-Action models, which take in visual input from cameras and language commands from a human operator, then output motor actions for the robot’s body. The idea is that a single end-to-end model could enable a robot to hear “pick up the red cup and put it on the table,” see the cup through its cameras, and execute the grasp and placement without separate perception, planning, and control modules.4National Science Review. Development of intelligent robots in the wave of embodied intelligence Chinese research groups and companies are pursuing this approach aggressively, often building on open-source large model architectures and fine-tuning them with robot-specific data.
The gap between demonstration and reliable deployment remains large. Current VLA models can perform impressive one-off tasks in controlled lab settings, like twisting a bottle cap or sorting objects, but struggle with the variability of real homes and factories. A robot that can open one specific cabinet in a lab may fail with a slightly different handle or a different lighting condition. Chinese companies are addressing this partly through brute-force data collection and partly through simulation-based data generation, which feeds into the digital twin work described below.
Digital Twins for Generating Training Data
One of the biggest bottlenecks for embodied AI is data. Unlike language models, which can train on text scraped from the internet, robot control models need physical interaction data: records of a robot reaching, grasping, walking, and recovering from mistakes. Collecting this data on real hardware is slow and expensive.
A team behind the humanoid robot Loong developed DT-Loong, a digital twin system that creates a high-fidelity virtual replica of the physical robot inside a simulation environment. By combining optical motion capture of human movements with algorithms that re-target those motions onto the robot’s body, DT-Loong can generate large volumes of training data for embodied AI models without requiring the real robot to physically perform each action.5PubMed Central. DT-Loong: A Digital Twin Simulation Framework for Scalable Data Collection and Training of Humanoid Robots The Loong robot itself uses a biomimetic joint design intended to mimic human movement more closely than conventional industrial actuators, which helps the motion re-targeting produce more natural results.
The digital twin approach addresses a specific frustration in humanoid development: even when the AI algorithms are ready, there is never enough good data to train them on. Several Chinese humanoid companies have invested in similar simulation infrastructure, often building on top of open-source physics engines and rendering tools. The general direction is toward “data factories” where thousands of virtual robots can train simultaneously in parallel simulated environments, compressing months of real-world experience into hours of computation.
Making Falls Less Dangerous
Any robot that walks on two legs will eventually fall. For humanoid robots working near people, the consequences of an uncontrolled fall could be serious for both the robot and anyone nearby. This makes fall safety a practical engineering priority, not just an academic curiosity.
A research team developed a unified fall-safety policy for the Unitree G1 that combines three behaviors into a single control system: preventing falls before they happen, reducing impact forces when a fall is unavoidable, and recovering to a standing position quickly after hitting the ground. The approach fuses a small number of human demonstrations of safe falling with reinforcement learning and an adaptive memory of safe reactions, and was tested both in simulation and on the real G1 robot.6arXiv. Unified Humanoid Fall-Safety Policy from a Few Demonstrations The results showed that the policy transferred from simulation to the physical robot and consistently produced lower impact forces and faster recovery across a range of disturbances, like being pushed or tripped.
What makes this line of work interesting is that it treats fall safety as something the robot actively manages, rather than something engineers try to prevent entirely through conservative movement. A robot that never takes a step that could lead to a fall is a robot that moves too cautiously to be useful. The more realistic goal is a robot that can handle the unexpected and get back up, much the way humans do. For commercial deployment in homes or factories, this kind of robustness is arguably more important than flashy locomotion demos on flat floors.
The Energy and Power Problem
Humanoid robots are power-hungry machines. Every joint motor draws energy, and the total consumption determines how long the robot can operate before needing a recharge. Battery life remains one of the most stubborn practical constraints. Most current Chinese humanoid robots can operate for somewhere between one and two hours of active movement on a single charge, depending on what they are doing.
A detailed study modeled the electrical power consumption of the Unitree G1’s arm across hundreds of different movement trajectories and speeds. The model captured how power draw varies across the arm’s joints with high accuracy, and revealed that different joints lose energy in distinct ways. Some joints are dominated by friction losses as they move, others by resistive heating in the motors, and one joint showed an unusual pattern where static friction was the primary energy drain.7arXiv. Identification of a Physics-Based Electrical Power Consumption Model for the Unitree G1 Humanoid Arm This kind of granular understanding matters because it tells engineers exactly where energy is being wasted and where design changes or software optimizations could extend battery life.
The energy challenge is not unique to Chinese robots, but it takes on particular importance given the commercial deployment timelines Chinese companies and policymakers have set. A robot that needs to recharge every 90 minutes is not yet practical for a full factory shift. Solutions under exploration include more efficient actuator designs, lighter structural materials, and software-level energy management that plans movements to minimize wasted power. Some companies have explored swappable battery packs, borrowing a concept from the electric scooter industry, to allow near-continuous operation with quick battery changes instead of long charging pauses.
Where Chinese Humanoid Robots Are Actually Showing Up
Despite the impressive demonstrations, most Chinese humanoid robots are still closer to advanced prototypes than mass-market products. The places where they are appearing today tend to fall into a few categories. Research labs and universities are the largest customer base, using platforms like the Unitree G1 and RobotEra’s XBot series as testbeds for AI and control research. Manufacturing companies, particularly in the automotive sector, have begun pilot programs where humanoid robots perform inspection, parts handling, or simple assembly tasks alongside human workers. And a small number of service-sector deployments have put humanoid robots in hotel lobbies, exhibition halls, and retail spaces, mostly in greeting and information roles rather than physically demanding work.
The factory use case is the one that Chinese companies talk about most. China’s manufacturing sector is enormous and faces growing labor shortages in certain regions and for certain types of repetitive work. A humanoid that can navigate a factory floor, pick up irregularly shaped parts, and place them where they need to go would be valuable even if it worked slowly, as long as it could operate reliably across a shift. Getting to that level of reliability is the current bottleneck, not the hardware itself.
Public perception of humanoid robots in general is more positive than many people assume. Survey research across different countries has found that people tend to hold moderately favorable attitudes toward humanoid robots, though with notable concerns about specific applications like law enforcement or caregiving. Gender differences in acceptance have been documented, with some studies finding that men and women differ in their comfort with robots in care and justice settings.8Technological Forecasting and Social Change. Public perceptions and acceptance of artificial intelligence humanoid bots/robots: Evidence from Turkey Within China specifically, the cultural reception has been shaped by the country’s long history of enthusiasm for technology adoption and by government messaging that frames humanoid robots as symbols of national technological progress.
How Chinese Humanoid Robots Compare Globally
The global humanoid robot field has a handful of prominent players outside China, and the comparison is instructive. Tesla’s Optimus project has attracted massive public attention but has been slower to reach independent real-world operation than some Chinese competitors. Boston Dynamics’ Atlas, long the most physically capable humanoid in the world, transitioned from a hydraulic research platform to an electric commercial design, but has not been deployed at scale. Figure AI and Apptronik are well-funded American startups pursuing similar goals.
Where Chinese companies have a potential edge is in manufacturing cost and speed of iteration. China’s supply chain for motors, sensors, batteries, and structural components is deep and competitive, which means a Chinese humanoid robot company can prototype and revise hardware faster and at lower cost than a counterpart sourcing components internationally. Unitree, for example, priced the G1 at a level significantly below what competitors charge for comparable platforms, a move that expanded access to the hardware and contributed to its widespread adoption as a research tool.
The areas where the global competition is most intense are in AI software and in the actuators that serve as the robot’s muscles. Chinese teams are competitive in reinforcement learning for locomotion and are investing heavily in VLA models, but much of the foundational AI model work still builds on architectures developed by American and European labs. On the hardware side, the actuators used in humanoid joints are a critical differentiator. Several Chinese companies have developed their own quasi-direct-drive and harmonic-drive actuators, aiming for the combination of high torque, low weight, and precise control that humanoid applications demand.
Dexterous Hands and Manipulation
Walking gets the headlines, but manipulation is arguably the harder unsolved problem. A humanoid robot that can stride across a room but cannot reliably pick up a coffee mug is not especially useful. Chinese research groups are actively working on dexterous robotic hands that can grasp objects of varying shapes, sizes, and fragility.
The challenge breaks down into two parts. First, the mechanical design of the hand itself: how many fingers, how many joints per finger, what kind of actuators, and how to fit enough dexterity into a hand-sized package without making it prohibitively heavy or expensive. Second, the perception and control side: how the robot senses what it is touching and adjusts its grip accordingly. Tactile sensors embedded in robotic fingertips can provide pressure and contact information, but integrating that sensory data into fast, reliable grasping behavior remains an open research problem.
Several Chinese companies have released multi-fingered hands designed to mount on their humanoid platforms. These typically have four or five fingers with multiple degrees of freedom each, and some include basic tactile sensing. The current generation can handle structured tasks like picking up boxes or tools from known positions, but struggles with the kind of improvised manipulation humans do effortlessly, like fishing a key out of a pocket or folding a towel. Closing that gap will likely require advances in both the AI models that plan grasps and the sensor hardware that provides feedback during contact.
The Role of Open-Source and Academic Collaboration
One distinctive feature of the Chinese humanoid robot ecosystem is a relatively open approach to sharing research tools and frameworks. Humanoid-Gym, for instance, was released as an open-source project, making the reinforcement learning pipeline accessible to any research group with a compatible robot and a GPU cluster.9arXiv. Humanoid-Gym: Reinforcement Learning for Humanoid Robot with Zero-Shot Sim2Real Transfer Similarly, the DT-Loong digital twin framework was published with enough detail to allow replication.10PubMed Central. DT-Loong: A Digital Twin Simulation Framework for Scalable Data Collection and Training of Humanoid Robots This pattern of building on shared infrastructure accelerates the field as a whole, even as individual companies compete on the commercial side.
Chinese universities play a significant role as both research contributors and talent pipelines. Tsinghua University, Peking University, Zhejiang University, and Harbin Institute of Technology are among the institutions with active humanoid robotics groups. Many of the startups in the space were founded by or employ researchers from these programs. The government’s policy push has also created dedicated funding streams for humanoid-related academic research, including grants specifically for sim-to-real transfer, embodied AI, and human-robot interaction.
The interplay between academia, startups, and government creates a feedback loop where research insights move to commercial products quickly, and commercial hardware makes its way back into labs as affordable research platforms. Whether this ecosystem will produce the first genuinely useful general-purpose humanoid remains to be seen, but the infrastructure for rapid iteration is firmly in place.