Kepler, Unitree and Fourier are showing a different side of China’s humanoid robotics push: productized platforms are becoming as important as headline-grabbing demos. Each company is making humanoids available, or at least commercially positioned, for developers, researchers, industrial partners and early adopters. That matters because broader access can accelerate experimentation, data collection and ecosystem formation, even if it does not yet prove that the robots are ready for repeat customer deployment.
The confirmed signal is that these companies are packaging humanoid robots as products rather than only research projects. Unitree lists low-cost humanoids with published pricing and open control interfaces. Fourier markets the GRx series around research, development tools and institutional users. Kepler has positioned its Forerunner K2 for industrial, logistics, inspection and research scenarios, while saying the robot has been tested at customer locations.
What remains unproven is commercial maturity. Product availability does not answer the hardest questions in humanoid robotics: uptime, safety, autonomy, service cost, intervention rate, real customer productivity, repeat orders and whether the robot can outperform cheaper non-humanoid automation. In Humanoid Analytics terms, these platforms are important technology signals, but they should not be treated as deployment proof.
China Is Turning Humanoids Into Platforms
The most important shift is not that Chinese humanoids can walk, dance or appear at trade shows. It is that several companies are trying to make humanoids accessible enough for third parties to build on them.
Unitree is the clearest example. Its R1 page lists a starting price of $4,900 for R1 AIR and $5,900 for R1, excluding tax and shipping, and describes the robot as an ultra-lightweight platform with open control interfaces for joints and sensors and support for simulation platforms. Unitree’s G1-D page goes further, positioning the platform around data acquisition, distributed training, model development and deployment support for open-source frameworks.
That is strategically important. A low-cost humanoid does not need to be commercially useful in a factory on day one to affect the market. It can still become a developer machine, a university research tool, a teleoperation testbed, a data collection device, or a way for integrators to learn what humanoid deployment actually requires.
Fourier is pursuing a higher-end but similarly platform-oriented path. Its GR-2 page describes the robot as “more agile, more powerful, more open,” and highlights customers or institutional users including ETH Zurich, Carnegie Mellon University, Simon Fraser University and Tsinghua University. Fourier lists GR-2 at 175 centimeters, 63 kilograms, with a two-hour battery, 53 joints and 12-degree-of-freedom dexterous hands. It also says its software development kit supports NVIDIA Isaac Lab, ROS and MuJoCo.
This is not the same as commercial deployment. But it is meaningful for research diffusion. Humanoid robotics needs more real hardware in more labs and developer programs because many of the sector’s bottlenecks, manipulation, energy use, teleoperation, sim-to-real transfer, balance recovery, safety and maintenance, cannot be solved by software demonstrations alone.
Kepler is positioning closer to industrial application. In a 2024 announcement distributed through PR Newswire, Shanghai Kepler Robotics said its Forerunner K2 had launched at GITEX Global and had been refined through engagement with nearly 50 target customers. The company said the robot was intended for intelligent manufacturing, warehousing and logistics, high-risk operations, and research and education, and that it had entered testing at customer locations on tasks including material handling, sample processing, patrol and inspection, stamping material collection and quality inspection.
Those claims are commercially relevant, but they are still mostly company-sourced. Kepler has not publicly provided enough independent customer confirmation, fleet size, uptime data, task economics or repeat-order evidence to classify K2 as operationally deployed at scale.
| Company | Platform Signal | What It Could Enable | Evidence Caution |
|---|---|---|---|
| Unitree | R1 and G1-D are positioned with pricing, open interfaces, data tools and model-training workflows | Developer access, university research, teleoperation, data collection and algorithm testing | Low price and platform access do not prove industrial autonomy or customer ROI |
| Fourier | GR-2 is marketed to research institutes with SDK support for Isaac Lab, ROS and MuJoCo | Embodied AI research, manipulation testing, simulation-to-real workflows and dexterity experiments | Institutional adoption is not the same as repeat commercial deployment |
| Kepler | Forerunner K2 is positioned for manufacturing, logistics, inspection and high-risk work | Early industrial trials, materials handling, inspection and customer-site learning | Customer testing claims need independent confirmation and operating metrics |
Lower Cost Can Change The Development Cycle
The pricing signal from Unitree is especially important because humanoid robotics has historically been constrained by hardware scarcity. If robots cost hundreds of thousands of dollars, only a small group of labs and well-funded companies can experiment with them. If entry-level humanoids fall into the thousands or low tens of thousands of dollars, many more teams can test controls, perception, teleoperation and embodied AI workflows.
That could change the pace of learning. More robots in more hands means more edge cases, more failures, more software tooling, more open-source experimentation and more developers who understand the practical limits of humanoid hardware.
Reuters reported in March 2026 that Unitree filed for a Shanghai IPO seeking to raise 4.2 billion yuan, about $610 million, and said the company’s prospectus showed humanoids had become a major growth driver. The same report said Unitree shipped more than 5,500 units in 2025, while noting that real-world factory deployment remained limited and that industry-application revenue mainly came from enterprise reception, tour-guide use, intelligent manufacturing and inspection, with tour-guide use accounting for roughly 50 percent to 70 percent.
That mix is exactly why evidence discipline matters. Unitree’s volume and pricing are important market signals. They also show that many current applications may still be closer to research, demonstration, reception, education or light inspection than high-value industrial labor replacement.
Productized Does Not Mean Production-Proven
Productization is a necessary step toward commercial maturity, but it is not the final step.
A robot can be available for sale and still be immature as an automation product. It may require expert operators, controlled environments, frequent maintenance, limited duty cycles or heavy teleoperation. It may be useful for labs but not yet robust enough for factories. It may perform impressive movements while still struggling with real customer tasks that require dexterity, judgment, safety and recovery from failure.
Fourier’s GR-2 illustrates both sides. The robot’s 12-degree-of-freedom hands, tactile sensing claims and support for common robotics development frameworks make it useful for embodied AI and manipulation research. But the public evidence does not yet show the kind of customer-site operating data that would establish deployment maturity. The stronger interpretation is that Fourier is helping build the research and developer layer of the humanoid stack.
Kepler’s K2 is similar but more industrially framed. The company says it engaged with customer requirements and tested the robot in real customer locations. That is stronger than a generic demo claim, but still not enough to establish paid pilot status or operational deployment without named customers, operating hours, robot counts, task results and evidence of customer continuation.
The Humanoid Analytics evidence standard is useful here: demos, shipment claims and product pages are technology or market signals, not proof of commercial readiness. The strongest evidence still comes from named customer pilots, paid pilots, operational deployments and repeat deployments.
The Ecosystem Effect May Matter First
The near-term value of Kepler, Unitree and Fourier may not be that they immediately replace workers. It may be that they lower the cost of learning.
A wider hardware base can help universities study locomotion and manipulation. It can help startups build software on real bodies rather than only in simulation. It can help integrators understand where humanoids fail in warehouses, factories and inspection settings. It can also help suppliers learn which actuators, batteries, hands, sensors and compute modules survive actual use.
That ecosystem effect is one reason China’s humanoid sector should be watched carefully. The country already has dense electronics supply chains, component suppliers, manufacturing capacity and state support for embodied AI. Productized humanoid platforms can turn those advantages into a feedback loop: more robots, more developers, more data, more component iteration and faster cost reduction.
The risk is that the market confuses ecosystem scale with deployment proof. A robot sold to a university, developer, demo venue or entertainment use case is not equivalent to a robot performing paid work in a factory or warehouse. A low-cost platform can be strategically important while still commercially immature.
The Next Proof Is Operational, Not Promotional
Kepler, Unitree and Fourier show that China is not only chasing humanoid spectacle. It is building product lines, software toolchains, developer access and research platforms that could expand the pace of experimentation. That is a meaningful technology signal for the global market.
But the standard for commercial proof remains unchanged. The next evidence to watch is named customer usage, paid pilot terms, operating hours, task completion rates, intervention frequency, maintenance cost, safety performance and repeat orders. For research platforms, the key proof is whether developers build durable applications on top of them. For industrial platforms, the proof is whether customers keep using them after the initial trial.
Product availability can accelerate the humanoid robotics learning curve. It cannot by itself prove that humanoid robots are ready for broad adoption. Kepler, Unitree and Fourier may help expand the ecosystem, but deployment maturity will still be decided in customer environments, not on product pages.
Sources:
Unitree, “Unitree R1”:
https://www.unitree.com/R1/
Unitree, “G1-D End-to-End Platform for Humanoid Robot”:
https://www.unitree.com/mobile/G1-D/
Reuters, “Unitree Plans Shanghai IPO, Testing Interest In Humanoid Robots”:
https://www.reuters.com/world/asia-pacific/unitree-plans-shanghai-ipo-testing-interest-humanoid-robots-2026-03-20/
Fourier Robotics, “GR-2”:
https://www.fftai.com/products-gr2
Fourier Robotics, “Fourier Robotics”:
https://www.fftai.com/
PR Newswire, “Kepler Debuts Forerunner K2 Humanoid Robot, Accelerating Commercial Deployment”:
https://www.prnewswire.com/news-releases/kepler-debuts-forerunner-k2-humanoid-robot-accelerating-commercial-deployment-302281546.html
Kepler, “Forerunner Series General-Purpose Humanoid Robot”:
https://www.gotokepler.com/productDetail
Humanoid Analytics, “Humanoid Company Tracker”:
https://humanoidanalytics.com/humanoid-company-tracker/
Related: The Humanoid Robot Market Is Splitting Between Evidence And Hype.