Unitree Rehearsal System Lets Humanoid Robots Teach Themselves

The robotics industry reached a pivotal inflection point this week as Unitree Robotics, China’s largest humanoid robot manufacturer, unveiled a self-improving AI system called Unitree Rehearsal. The system, announced by founder Wang Xingxing at a physical AI conference in Hangzhou, enables humanoid robots to autonomously learn, write code, test themselves in simulation, and refine their behavior through real-world feedback loops. The breakthrough represents a fundamental shift in how robots acquire skills, moving from human-programmed instructions to self-directed learning pipelines.

How Unitree Rehearsal Works

The Unitree Rehearsal system operates on a multi-stage learning architecture. First, the AI autonomously searches academic research papers to identify new techniques and methodologies relevant to its task domain. It then writes code implementing those techniques and tests the code in a simulation environment before deploying it on physical robots. Real-world trial data is collected and fed back into the model to adjust parameters and improve performance iteratively.

This approach mirrors the reinforcement learning paradigm that has driven recent advances in large language models, but applied to the physical world. Wang described Rehearsal as the company’s most important project of the year, signaling that self-improvement capabilities are now seen as the critical path to practical humanoid deployment.

• Autonomous research: The AI scans and synthesizes academic literature without human intervention
• Code generation: New behaviors are programmed by the AI itself, tested in simulation first
• Real-world feedback: Physical robot trials generate data that refines the underlying models
• Iterative improvement: Each cycle produces a more capable robot without manual reprogramming

The 2028 Breakthrough Timeline

Wang Xingxing made headlines by dramatically accelerating his timeline for humanoid robot practicality. He now predicts that humanoids will complete 80% of assigned tasks in factory or home settings through simple voice commands as soon as 2028. He defines this threshold as the “aha moment” for practical robotic applications, the point where robots become genuinely useful to ordinary consumers and businesses.

This prediction represents a significant reversal for the 36-year-old entrepreneur. At the World Robot Conference in Beijing this past August, Wang suggested that mass deployment in factories and homes could take up to a decade. He credited recent advances in AI models, particularly OpenAI’s GPT-6, with changing his assessment. Wang admitted he had been skeptical about AI model developers’ progress last year, but the emergence of self-improving frontier models convinced him the timeline was shorter than previously believed.

“As long as you give a frontier model data and time, it will undoubtedly surpass humans in many areas.”

Wang was candid about current limitations, however. He acknowledged that if you bring a humanoid robot home today and ask it to clean a room, it honestly cannot complete that task reliably. The gap between today’s capabilities and the 2028 vision is precisely what the Rehearsal system is designed to close.

H2-X: The Robot Bodyguard

Alongside the Rehearsal announcement, Unitree posted a dramatic video of its H2-X Kung Fu Legend Guardian, a humanoid robot designed for personal protection. The video shows the robot punching and kneeing cars in a scrapyard, confronting an intruder wielding a knife, and lifting a vehicle to free a trapped person. Unitree claims the robot operates for approximately eight hours on a single charge and can jump from a four-meter platform without damage.

The H2-X builds on the company’s H2 platform, which retails for $29,900. Unitree has not announced pricing or a release date for the H2-X variant. Industry observers have noted that the robot would likely face significant safety certification challenges before it could be sold as a personal bodyguard. The video itself carries a warning label reading “extreme testing” and advises against imitating the depicted scenarios.

Scaling Production for Millions of Robots

Unitree currently manufactures tens of thousands of humanoid robots annually. Wang indicated the company plans to expand production capacity dramatically to meet potential demand for tens of millions of robots as capabilities improve and prices decline. This ambition aligns with broader industry trends, as vendors shipped approximately 25,000 humanoid robots in the first half of 2026 alone, already exceeding the roughly 18,500 shipped in all of 2025.

The company is pursuing a dual-track AI strategy, building its own proprietary models while simultaneously collaborating with AI laboratories in China and abroad. Unitree’s humanoids running GPT-6 can already paint with changing colors and palette rinsing, and play the board game Go, demonstrating increasingly sophisticated fine motor control and strategic reasoning capabilities.

Market Context and Investment Landscape

Unitree’s shares rose fivefold on their Shanghai STAR Market debut in August 2026, making it the first humanoid robot stock on China’s A-shares market. Shares have since retreated to approximately a quarter below their August peak, though the company remains one of very few in the humanoid space to sustain profitability across consecutive years. Chinese venture and state funds have poured more than 100 billion yuan, approximately $15 billion, into robotics startups this year alone, according to tracker ITjuzi.

However, China has reportedly slowed new humanoid robot IPOs since Unitree’s listing, suggesting regulators are taking a more measured approach to the sector’s rapid expansion. For now, humanoids sell primarily to researchers and enthusiasts rather than mass-market consumers, a reality Wang openly acknowledges even as he accelerates his timeline for mainstream adoption.

The Competitive Global Race

Unitree’s announcement comes amid intensifying global competition in humanoid robotics. In the United States, Agility Robotics recently expanded its collaboration with NVIDIA as it prepares its Digit robot for commercial-scale manufacturing deployment. AGCO Corporation signed a commercial agreement with Agility Robotics to pilot the Digit robot in its Hesston, Kansas manufacturing facility beginning in 2027, targeting repetitive and hazardous assembly tasks. Meanwhile, new US regulatory restrictions on certain robotics technologies have created headwinds for Chinese manufacturers seeking American market access, while potentially accelerating European adoption.

The convergence of self-improving AI systems like Rehearsal with increasingly capable humanoid hardware suggests the industry may indeed be approaching its ChatGPT moment. Whether that breakthrough arrives in 2028 as Wang predicts or takes somewhat longer, the trajectory is unmistakable. Robots that can teach themselves represent a qualitatively different category of machine, one that improves autonomously rather than waiting for human engineers to program each new capability.

Implications for Industries and Consumers

For manufacturers, the prospect of self-improving humanoids could dramatically lower the total cost of ownership. Rather than requiring expensive reprogramming when production lines change, robots equipped with Rehearsal-style systems could adapt on their own. For consumers, the 80% task completion threshold Wang identifies would represent a transformative shift from novelty to genuine utility in household settings.

The broader implications extend beyond economics. Self-improving robots raise important questions about safety, oversight, and the rate at which autonomous systems should be permitted to modify their own behavior. As Unitree and its competitors push toward the 2028 milestone, these governance challenges will become increasingly urgent for regulators, manufacturers, and consumers alike.

Stanford University researchers also recently released a free online textbook and curriculum covering autonomous robotics, reflecting growing recognition that education and workforce preparation must keep pace with rapid technological advancement. The democratization of robotics knowledge may prove as important to the field’s future as the hardware and AI breakthroughs generating headlines.

---

Edited by Palawan @QUE.COM
Website: https://QUE.COM Intelligence
Sponsored by: https://MAJ.COM AI Autonomous