Unitree Humanoid Robots: The Road to Self-Improving AI
Unitree is building AI that helps humanoid robots improve through research, simulation and testing. See what its 2028 forecast means for business and safety.
Unitree's Self-Improving Robots Could Change the Humanoid AI Race
Unitree Robotics founder Wang Xingxing has laid out a pretty bold forecast for humanoid robots. He expects them to complete 80% of voice-given tasks by 2028. Unitree is also working on a system called Unitree Rehearsal, aimed at helping robots learn from research, code, simulation, and tests in the physical world.
On the surface, these are just claims about smarter machines. But the bigger idea is about shrinking the gap between a model's proposal and a robot's real-world performance - so improvements cycle faster in a factory or at home. If that loop works reliably, it could change how companies train, roll out, and update robots. Still, predictions aren't benchmarks, and demos aren't the same as proof that a robot can operate safely and independently in everyday conditions.
The useful question isn't whether humanoids will replace people overnight. It's where a faster improvement loop could create real value, what still needs to be proven, and how businesses should treat timelines that sound almost too good to be true. Here's what Unitree says it's building, what Wang's 2028 prediction appears to mean, what the H2-X video does - and doesn't - show, and the practical constraints that come with selling anything like this at scale.
What Unitree Rehearsal Is Meant to Do
Unitree's self-improvement system is called Unitree Rehearsal. Wang Xingxing described it as the company's most important project of the year at a physical AI conference in Hangzhou, according to a report by the South China Morning Post.
As described, the workflow ties together stages that are often handled separately. It uses AI to search research, write code, test changes in simulation, and then move promising results to trials with real robots. Feedback from those physical tests can then feed back into the next set of model updates.
That matters because robots don't "see" the world the way software models do. A system can generate an answer that sounds right while still failing on the messy details: grip, weight, balance, friction, lighting, and the fact that objects are rarely exactly where the system expects them. Simulation helps you test lots of options quickly, but the physical world still has the final say.
The value is in the feedback loop
A typical robotics development cycle can be slow. Engineers find a failure, reproduce it, adjust the system, test the change, and then check whether it broke something else. That's necessary - especially when real hardware is involved. But it can get painful when a robot has to deal with lots of different objects and environments.
Automation can help parts of the process. A model could pull in relevant research, propose code changes, or run repeated simulated trials before engineers spend time testing on a physical robot. The key benefit, at least in theory, isn't that it eliminates people from development. It's that it can help teams explore more candidate fixes using the same engineering time.
There's also a hard limit: more trials don't automatically produce better robots. Simulation can be incomplete. A model can optimize for something that looks good in testing but doesn't match the real job. And generated code still needs review, validation, and rollback plans before it's safe to control equipment near people.
What "self-improving" should mean in practice
"Self-improving" can sound like a robot quietly rewriting its own intelligence. What Unitree describes is more specific: using AI to support research, coding, and testing, with real-world feedback used to update models. That's an important difference. Speeding up a supervised improvement pipeline isn't the same thing as an autonomous system that can change its own operating rules safely without oversight.
For businesses looking at physical AI, the practical measures are straightforward. How often does the robot complete the task without intervention? How does it behave when the object, workspace, or instructions change? Can a team trace what changed between software versions? And if a new update performs worse, can it roll back to something known to work?
Those questions are less flashy than "machines improving themselves," but they're what determine whether the system is useful outside a demo.
What the 2028 Prediction Actually Says
Wang expects humanoid robots to complete 80% of voice-given tasks by 2028, according to Bloomberg's account of a transcript shared by Unitree. The forecast is framed around tasks in homes or factories. Wang reportedly called the moment when robots reach that level the "aha" point for practical humanoids.
The number sounds concrete, but it needs context. The source doesn't provide a standardized test set, task definitions, failure costs, or how much human help counts as "completion." "Complete 80% of tasks" could mean success on a carefully chosen set of short activities. It could also mean completion under specific conditions, with a human standing by to intervene. Those are very different levels of capability from reliably handling a live, open-ended job on a real production line.
Wang also offered a qualification of his own. In the reported transcript, he said that if someone brought a humanoid home today and asked it to clean a room, he didn't think it could do the task yet. That's exactly where robotics tends to get difficult: bridging the gap between narrow demonstrations and everyday household work.
A benchmark needs more than a percentage
If someone claims a robot can "complete 80%," you want at least a few details:
- The task set: Are these repetitive, controlled tasks, or varied, open-ended ones?
- The operating conditions: Is the robot working in a prepared environment, or one that changes from day to day?
- The success threshold: Does partial completion count? How are errors scored?
- The level of supervision: Can a person intervene, reset the robot, or guide it through the task?
Without that, the 80% figure is best read as a directional forecast, not something that proves readiness for deployment. It does tell you how quickly Unitree believes progress is happening, but it doesn't answer the purchasing question on its own.
Why Wang says his outlook changed
Wang said advances in AI models changed his view. He reportedly described being skeptical of AI model developers the previous year, then later seeing leading developers allow models to improve themselves. He also said that, with data and time, a frontier model would "undoubtedly surpass humans in many areas."
That's a broad statement about AI progress. It doesn't automatically mean humanoid robots will beat humans at general physical work. Language and coding models can improve in digital environments without necessarily gaining reliable movement, perception, or judgment in the real world. A robot still has to turn instructions into actions, detect when the situation has changed, and avoid causing harm.
Unitree says its humanoids running GPT-6 can paint (including changing colors and rinsing a palette) and play Go. Those are examples the reporting cites, not independent proof of general-purpose competence. They do show the kinds of tasks teams are using to combine models with physical hardware. The tougher test is whether the robot can repeat useful work across ordinary variations, at a cost and risk profile that makes sense.
The H2-X Video and the Gap Between Demos and Deployment
Unitree also published a one-minute video featuring the H2-X Kung Fu Legend Guardian. The footage shows the humanoid punching and kneeing cars in a scrapyard, confronting an intruder holding a knife, and lifting a car to free someone trapped underneath. Unitree says the robot can run for about eight hours. TechRadar reports that it can jump from a four-metre platform.
These scenes are clearly meant to show power and movement. They shouldn't be treated as evidence that the robot is ready to act as a personal bodyguard or emergency responder. A staged demonstration can show a machine performing a movement in specific conditions, but it doesn't, by itself, establish reliability in unpredictable situations, safe decision-making, or compliance with requirements for commercial use.
The weapon scenario raises the stakes even further. A robot that responds incorrectly could escalate a situation or injure someone nearby. A demonstration of physical force doesn't answer the real questions: who authorizes the robot to act, how it distinguishes a genuine threat from a misunderstanding, and who takes responsibility if its decision causes harm.
Safety is an operating requirement
The video labels the scenes "extreme testing" and warns viewers not to imitate them, according to Gizmodo's report. Unitree's post also asks people to use the robot safely and in a friendly manner. Those warnings are reasonable - but they aren't the same as a safety case.
Before a humanoid could be sold and used around the public, it would need to meet relevant safety certification requirements. TechRadar notes that the H2-X would struggle to pass the certification needed for that kind of role. The reporting also says Unitree hasn't announced a price or release date for the H2-X. The existing H2 model is listed at $29,900, but that number shouldn't be assumed to apply to the more specialized machine.
For any robot meant to operate near people, safeguards usually need to include controlled operating zones, clear emergency stops, limits on force and speed, human override, logging, and a way to disable the system quickly. The exact requirements depend on the task and jurisdiction - warehouse robots, home assistants, and security products don't carry the same risk profile.
Demonstrations are useful - but narrow
Video demos can help you understand what a machine is designed to do, and they also reveal what a company wants to emphasize. Here, the H2-X video foregrounds strength, mobility, and dramatic intervention. A buyer evaluating a real deployment would need different evidence: repeatable task results, failure rates, maintenance needs, safe behavior around people, and the cost of supervision.
That's not a reason to dismiss the demo. It's a reason to read it carefully. A capable movement is a real technical milestone. But a product is a full system: it has to handle edge cases, meet standards, be maintained, and deliver enough value to justify its price.
The Business Case for Humanoid Robots
Unitree's ambitions are landing in a market that's drawing major investment and equally major hype. The company reportedly makes tens of thousands of robots per year and plans to expand production in anticipation of demand for tens of millions. Bloomberg also reports that Chinese venture and state funds invested more than 100 billion yuan - about $15 billion - in robotics startups this year, citing ITjuzi.
Those numbers suggest strong interest, not proof that a mass market is already here. The reporting notes that humanoids currently sell mostly to researchers and enthusiasts. The gap between those early buyers and broad use in factories or homes is still large. Research customers can tolerate limited capability because experimentation is part of what they're buying. A factory operator, by contrast, wants a robot that performs a defined job, fits into existing workflows, and can be supported when it inevitably breaks.
Unitree's shares rose five-fold on their Shanghai debut in August and were about a quarter below their August peak at the time of the report. China has also reportedly slowed humanoid IPOs. Share-price moves and funding levels reflect expectations about future markets, not confirmation that any particular deployment already works economically.
Start with the job, not the humanoid form
A humanoid shape can be useful when a robot has to work in spaces built for people - doors, tools, workstations, and storage that already exist. But copying human form isn't automatically the cheapest way to automate a task. If the work is repetitive and the environment can be designed around it, a dedicated arm, a mobile platform, or a fixed machine might be simpler and more reliable.
A business considering a humanoid should start with a specific task and measure things like:
- How often it happens and how much labor time it currently takes
- How much variability it involves, and what happens when the robot can't continue
- The cost of hardware, integration, maintenance, training, and human supervision
- Expected uptime and the financial impact of downtime or a safety incident
- Whether a simpler automation option could handle the same job more reliably
This helps avoid a common mistake: buying a general-purpose platform before you've identified a problem that genuinely justifies it. In real deployments, the best first use is often a narrow, measurable task with a reasonably controlled environment and a manageable cost of failure.
What would make the market scale
To move beyond researchers and enthusiasts, several parts of the system have to improve together. Better AI is only one piece. Robots also need dependable hardware, batteries that can support useful shifts, service networks, software updates that don't introduce regressions, and clear ways to supervise and recover from failures.
Unitree's plan to develop its own models while working with AI labs in China and abroad suggests it's thinking across the whole stack. That could give the company more control over how models connect to its hardware. But it also means it has to manage a broad engineering and operational challenge: robot, model, training process, and support process all need to work together.
A self-improvement loop might speed up learning, but it doesn't remove the need for careful testing. If an AI-generated change makes a robot faster at one task while making it less predictable around people, that's not an "improvement" from an operator's perspective. The real test isn't whether a capability can be demonstrated. It's whether the full system delivers reliable value at a total cost customers can defend.
Conclusion
Unitree's self-improvement plan and Wang's 2028 forecast capture the current ambition in physical AI: bring powerful models closer to machines that can act in the real world, then use those actions to improve the next version. If that improvement loop really holds up, it could shorten development cycles and make humanoids more useful in clearly defined settings.
But the gap between a forecast and a dependable product still comes down to task reliability, safety, serviceability, and economics. For businesses, the smart approach is to evaluate robots against a specific job, demand transparent test conditions, and compare them with simpler automation. For everyone else, it's worth treating impressive demos as progress - not proof that general-purpose humanoids are ready for everyday life.
The next milestone that would actually matter isn't another dramatic video. It's independently verifiable performance on a clearly defined task, repeated under realistic conditions, at a cost that makes deployment worth it.