The controversy around Unitree Robotics seems to be about whether robots can do flips, whether demos are remotely controlled, and whether an IPO is mainly an exit channel.
At a deeper level, it exposes a structural mismatch in China’s embodied AI industry: capital markets and public attention are already pricing a general-purpose robot era, while the industry is still in early stages of demos, pilots, data accumulation, and commercial-loop validation.
Unitree is not a story company without technology. It represents strong engineering, cost control, and communication capability in China’s robotics industry. The real question is not whether the technology exists, but whether it can support real productivity scenarios and repeat purchases.
Moving Is Not the Same as Understanding Tasks
For humanoid robots, motion is what the public sees first. Running, jumping, and flipping are naturally easier to spread, but motion capability and intelligence are not the same thing.
A general robot can usually be split into brain, cerebellum, and body: the brain handles cognition, task planning, and embodied foundation models; the cerebellum handles motion control, path planning, and gait balance; the body covers physical structure and actuators.
Complex movement first proves body design, low-level control, and engineering calibration. Whether the robot can understand environments, decompose tasks, handle exceptions, and keep learning depends on its brain. Moving is the prerequisite for industrialization; understanding tasks is where commercialization begins.
Selling Products Is Not the Same as Proving Productivity
One underappreciated point about Unitree is that it already sells products well. In 2025, Unitree reported revenue of 1.699 billion yuan and adjusted net profit of 591 million yuan, which is rare in an early humanoid robotics market.
But revenue scale does not mean humanoid robots have passed real market validation. The source of revenue matters. A large share still comes from research, education, developers, and demo-like scenarios, while manufacturing, logistics, inspection, and other productivity scenarios remain limited.
Research procurement proves that people will buy a robot platform. Industrial and service scenarios prove whether robots can keep creating productivity.
Only when robots operate steadily in production lines, warehouses, inspection, and service scenarios, and customers repurchase because efficiency improves, costs fall, and failure rates stay manageable, has the productivity logic truly worked.
Capital Can Price the Future, but Industry Cannot Pre-Cash It
The IPO debate also needs precision. Listing can improve financing capacity, brand credit, and governance, while giving early investors a future exit path. Embodied AI connects AI, robotics, manufacturing, labor substitution, and industrial policy, so early capital bets are expected.
The real issue is that capital markets can turn the future into today’s valuation too early. Robotics is not an internet product. It cannot validate value only through user growth, parameter scale, or viral attention.
It has to face hardware reliability, supply chains, delivery, after-sales service, operations, safety, failure rates, and customer ROI. Capital can accelerate research, but if it runs too far ahead, companies are pushed into telling a large-scale commercialization story too early. Industry cannot cash in the future before it arrives.
The Industry Has Not Yet Passed the Scale Test
If we zoom out from Unitree to China’s embodied AI industry, routes differ, but every company faces the same gate.
UBTech looks more like a full-stack, heavy-investment route with more industrial pilots and R&D. AgiBot emphasizes data, models, and scenario loops, and has started to show industrial orders. Galaxy General, X Square Robot, TARS, and similar players focus more on embodied models, VLA, world models, real-world data collection, and generalization.
These routes all matter, but they must answer the same questions: can model capability become stable delivery, is real-world data sufficient, will customers keep paying, and can revenue, margin, repurchase, operations cost, and failure rates survive long-term inspection?
The Real Competition Is the Data Loop
The first half of humanoid robotics is about who can build robots, move them, and sell them. The second half is about who can make robots keep learning, reduce costs, and create value in the real world.
The valuable flywheel should look like this:
- Low-cost bodies enter more real scenarios.
- Real scenarios produce high-quality behavior data.
- Data improves embodied models and task strategies.
- Better models enter more complex scenarios.
- Broader deployment raises shipment volume and lowers hardware cost.
- Lower cost improves customer ROI, repurchase, and deployment scale.
Unitree is strong in the first half of the loop: low-cost bodies, motion control, productization, manufacturing efficiency, and market attention. AgiBot, Galaxy General, X Square, and others emphasize the second half: data, models, generalization, and scenario intelligence. UBTech tries to connect the whole chain with heavier R&D and industrial delivery.
It is too early to declare which route wins. The industry is not in the final round yet; it is still competing for the right to enter it.
Final Thoughts
The value of the Unitree controversy is that it puts the industry’s contradictions on the table early: motion versus intelligence, capital heat versus commercial validation, research procurement versus real productivity, hardware engineering versus the AI brain, and short-term attention versus long-term repurchase.
China’s embodied AI industry does have opportunities: a complete hardware supply chain, rich manufacturing scenarios, fast engineering iteration, and strong cost-down capability. But these advantages must cross one threshold: moving from demo robots to productivity robots.
The winner will not simply be the company with the best video demos, the most funding, or the highest valuation. It will be the company that closes the commercial loop first.
The real second half of robotics is not making robots look human. It is making robots as reliable, affordable, and stable as tools that create value.