Wang Xingxing built robots that can do backflips, lift cars, and perform kung fu. At a conference in Hangzhou on October 22, he told the audience those skills are no longer the bottleneck. The bottleneck is that his robots still need human engineers to write their control code, and he has a plan to fix that. By 2028, he says, the company's humanoid robots should complete 80% of assigned tasks in factories and homes after receiving nothing more than verbal instructions. Two months ago he said the same milestone was ten years away.
What Actually Happened
Bloomberg reported on October 23, 2026, that Unitree founder Wang Xingxing delivered a forecast at the World Physics AI Conference in Hangzhou that put practical humanoid robotics two years away instead of a decade. The 80% threshold is his own definition of the "aha moment" for humanoid adoption: the point at which a robot can reliably execute four out of five requested physical tasks, after hearing those tasks described verbally, without pre-programming or demonstration. Wang attributed the compressed timeline to breakthroughs in large language model capability, specifically citing GPT-6-class models that can now guide robots through untrained physical jobs by breaking them down into motor primitives that the robot's control system can execute.
Alongside the forecast, Wang announced Unitree Rehearsal, which he described as the company's most important project of the year. Rehearsal is an AI feedback pipeline in which language models generate robot control code, test that code in simulation, run the best-performing variants on physical robots, score the outcomes, and feed the results back into the next generation of code. Humans remain in the loop to assess performance and provide qualitative feedback, but the system is designed to progressively automate the upgrade process as performance improves. Wang said the system specifically targets the "last few centimeters or millimeters" of precision that current humanoid robots consistently fail on: the final adjustment when grasping an object, placing a component, or completing a fine motor task. The gap between a robot that can lift a car and a robot that can reliably pick up a single egg has remained stubbornly persistent across the industry for years.
Unitree's competitive position is already strong by the numbers that are publicly available. The company shipped more than 7,000 humanoid units in the first half of 2026, second only to AgiBot's approximately 9,700 units in a global market that delivered roughly 22,000 humanoid units total in H1, according to industry estimates. The Next Web notes that Chinese manufacturers accounted for roughly 97% of global humanoid shipments in H1 2026, a concentration that underscores how thoroughly the competitive center of gravity has shifted toward Chinese robotics companies over the past 18 months. Unitree's H2-X robot, also revealed this week, is designed for security and rescue operations, with a claimed eight-hour operating time on a single charge.
Why This Matters More Than People Think
The timeline compression from "ten years" in August 2026 to "two years" in October 2026 is the most important number in Wang's announcement, and not primarily because of the 2028 target itself. What matters is the mechanism that drove the compression: a different class of AI model became available, and it changed the capability calculus for the entire industry almost instantly. Wang is describing GPT-6-class language models as a prerequisite for 2028 humanoid viability. That is a claim about dependency and architecture that has broad implications: the pace of humanoid robot development is now partially a function of frontier AI model development, which means the timeline for humanoid adoption is linked to the same training cycles and capability jumps that drive the LLM industry. The two trajectories are merging.
The Rehearsal system represents a specific and underappreciated structural shift in how robot capability will be developed going forward. The traditional robotics engineering pipeline requires human experts to write, test, and iterate control code manually, a process that is slow, expensive, and limited by the bandwidth of the engineering team. What Unitree is describing is the application of reinforcement learning from human feedback (RLHF), the same technique that produced the modern generation of helpful AI assistants, to physical robot control. If the feedback loop works as described, robot capability improves at a rate determined by simulation throughput rather than engineering headcount. Unitree has hundreds of robots in operation and can run thousands of simulated trials per hour. The team of human engineers providing feedback becomes a multiplier, not a bottleneck.
The 2028 forecast also resets the standard for what counts as a practical humanoid deployment. Wang's 80% threshold, completing four out of five verbally requested tasks, is a very specific claim. It implies that a homeowner could tell a Unitree robot to pick up laundry, load a dishwasher, carry groceries, and wipe a counter, and the robot would complete four of those five tasks correctly without any pre-programming or demonstration. Whether that level of performance is achievable by 2028 is genuinely uncertain. But the fact that the founder of the world's second-largest humanoid manufacturer is willing to commit publicly to that target in two years, having just called it a decade away, is the kind of inflection signal that the investor and enterprise community will be watching closely.
The Competitive Landscape
The backdrop to Wang's announcement is that Tesla started production of the Optimus robot on a high-volume line at its Fremont factory, as disclosed in the company's Q3 2026 earnings call on October 21. Elon Musk described it as "early production," noted that a 20,000-unit weekly target is about a year away, and said external customer deliveries would begin around the end of 2027. The timing of Wang's 2028 forecast, arriving two days after Tesla's production announcement, is not likely coincidental. Both companies are now on explicit timelines to commercial humanoid deployment, and the race between the Chinese manufacturers who lead in unit volume and Tesla, which leads in public attention, is moving into a phase where production ramp rates and task capability will be the competitive metrics.
Figure, the US humanoid startup, reported more than 1,250 hours of commercial operation at BMW's manufacturing facility with its Figure 03 model running the Helix 2.5 control system. Boston Dynamics Atlas, now several generations into its commercial development, published updated hand specifications in early October, with 13 degrees of freedom per hand. AgiBot, Unitree's closest Chinese competitor by unit volume, has not publicly announced a self-improving AI system comparable to Rehearsal, though several Chinese labs have been working on similar pipelines in research settings. The Rehearsal announcement, if it works as described, gives Unitree a potential long-term capability advantage over competitors still relying on manually written control code.
A useful historical parallel is the shift in software development from human-written test suites to AI-generated tests. For years, experienced engineers wrote unit tests by hand, a process that was slow and often incomplete. When AI coding tools became capable enough to generate test code automatically, the velocity of software development increased not because the AI was better than a good engineer, but because it could cover more ground faster and at lower cost. The robotics engineering pipeline faces the same structural constraint: the number of physical tasks that need to be programmed, tested, and refined is vastly larger than the engineering team available to handle them. Rehearsal, if it works, addresses that constraint with simulation throughput rather than hiring.
Hidden Insight: The Two-Month Reversal and What It Actually Means
Critics should not accept Wang's two-month timeline compression at face value. In August 2026, at the World Robot Conference in Beijing, Wang said humanoid robots would take ten years to replace human workers at scale. In October 2026, he says the 80% task-completion milestone is two years away. These statements are not strictly contradictory. "Replacing human workers at scale" and "completing 80% of tasks after verbal instructions" are different thresholds. But the gap between them is still sharp enough to invite skepticism. The bear case is that Wang made an aggressive public forecast in a competitive environment, shortly after Tesla's production announcement, knowing that timeline claims are a form of market positioning in the humanoid race. CEOs in competitive industries have strong incentives to pull timelines forward in public statements even when the internal engineering reality is more cautious.
The Rehearsal system's actual performance is also unverified. Wang described a pipeline architecture at a conference and said it targets the precision gap. He did not present benchmark results, video evidence of the system running on physical robots, or a technical paper. The science of robot learning through simulation has well-documented limitations: sim-to-real transfer, the process of taking behavior learned in simulation to physical hardware, fails in ways that are difficult to predict until the physical tests run. Unitree's own H2-X robot stumbled and fell during a side flip at the same conference where Wang was speaking, a visible reminder that the precision gap Wang is targeting is a deeply real engineering challenge, not just a software problem to be solved with better AI models.
The insight that most analysis of this announcement will miss is what the Rehearsal system implies about open-source robotics. Unitree has historically been more open than most hardware robotics companies: its quadruped control code has been available on GitHub, and the research community has built extensively on Unitree's hardware because of that openness. If Unitree publishes Rehearsal as an open platform, the training loop it describes would become available to every university robotics lab with Unitree hardware and a GPU cluster. The leverage effect of that would be enormous: instead of one company's engineering team improving robot capability through Rehearsal, the global robotics research community would be contributing to the same pipeline. Wang has not announced whether Rehearsal will be open-sourced, but given Unitree's track record, the question is worth asking explicitly.
The dependency on frontier AI models that Wang acknowledged is also a double-edged strategic position. If GPT-6-class models are enabling the capability jump he describes, Unitree's roadmap is partially dependent on the continued improvement of AI systems it does not control and cannot predict. A training policy shift like OpenAI's October 23 announcement, moving frontier training offline and potentially affecting the capability trajectory of future models, could affect Unitree's timeline in ways that Wang cannot account for in a two-year forecast. The robotics and AI ecosystems are now coupled tightly enough that a decision in San Francisco about how to train a language model ripples through to manufacturing floors in Hangzhou.
What to Watch Next
The 30-day test is whether Unitree releases any technical documentation for Rehearsal: a paper, a GitHub repository, benchmark results from simulation runs, or video evidence of the training loop improving physical robot performance. The conference announcement described an architecture. Technical credibility requires evidence. If documentation appears within a month, it suggests the system is real and at a stage where external validation is useful. If the 30-day window passes without a technical release, the announcement should be weighted as a roadmap claim rather than a demonstrated capability.
The 90-day indicator is the cleaning and household task demonstration that Wang described as a target for 2027. Unitree has a history of releasing compelling robot demonstration videos, from parkour runs to the H2-X martial arts performance. A video showing a Unitree humanoid completing household tasks after verbal instruction, without preprogramming, would be a concrete milestone that the market can evaluate and competitors will need to respond to. The gap between a demonstration video and commercial deployment is large, but a credible video changes the conversation.
The 180-day question is what Tesla Optimus's production ramp actually delivers by mid-2027. If Tesla is producing several thousand Optimus units per week by April 2027, as the roadmap with the AI5 chip integration implies, the competitive pressure on Unitree intensifies regardless of how well Rehearsal performs in simulation. Wang's 2028 forecast is implicitly a claim that Unitree's software capabilities will match or exceed Tesla's hardware scale. Whether both claims prove correct, or whether the commercial humanoid market in 2028 looks like a two-company race between the leading Chinese manufacturer and the leading American brand, is the most important question in physical AI right now.
Unitree's founder just compressed his own 10-year forecast to 2 years in two months. The reason he gave, frontier AI models enabling robot task generalization, is more important than the date.
Key Takeaways
- 80% task completion target, 2028: Unitree founder Wang Xingxing forecasts humanoids completing four of five verbally requested physical tasks by 2028, revised sharply from his August prediction of a 10-year timeline to scale
- Unitree Rehearsal announced: An AI feedback loop in which language models generate robot control code, test it in simulation, run physical trials, and iterate automatically, targeting the persistent "last few centimeters" precision gap
- 7,000-plus units shipped in H1 2026: Unitree is the world's second-largest humanoid manufacturer by unit volume, trailing only AgiBot's approximately 9,700 units in a market where Chinese manufacturers hold roughly 97% of global shipments
- GPT-6-class models cited as the enabler: Wang attributes the accelerated forecast directly to frontier AI models now capable of guiding robots through untrained physical tasks by decomposing verbal instructions into motor primitives
- H2-X security robot revealed: The "Kung Fu Legend Guardian" robot debuted for security and rescue applications, with a claimed eight-hour operating time, though it stumbled during a live demonstration at the same conference
Questions Worth Asking
- Wang compressed his timeline from 10 years to 2 years in two months based on a single generational jump in AI model capability. If that dependency holds, what happens to humanoid robot timelines if frontier model development slows or hits a capability plateau?
- The Rehearsal system trains robots by scoring simulation performance and iterating in a loop that could theoretically run continuously. What oversight mechanisms prevent a self-improving robot training loop from optimizing for measurable simulation metrics at the expense of real-world safety behaviors?
- Unitree ships the most humanoid robots outside Tesla by volume, and its hardware has been available to researchers globally. If Rehearsal is open-sourced, does that change the competitive dynamics of the humanoid race from hardware differentiation to software ecosystem, and who wins in that version of the competition?
