China's Robots Just Did Kung Fu on National Television. The West Is Behind.

On February 17, 2026, China's Spring Festival Gala aired on CCTV. It is the most-watched television broadcast in the world — nearly one billion viewers, every year. This year, alongside the fireworks and folk dances, Unitree's humanoid robots performed a fully autonomous kung fu and parkour routine in front of a live audience.

They did not stumble. They did not fall. They vaulted continuously, performed a 7.5-rotation Airflare grand spin, and completed an aerial flip reaching over three metres in the air. All of it autonomous. All of it in real time.

Twelve months ago, China's showcase robot performance featured a Unitree model carefully waving a handkerchief. The contrast is not gradual improvement. It is a step change.

What actually happened at the gala

The performance featured Unitree's G1 and H2 humanoid robots. Several world-firsts were claimed and verified:

• World's first continuous freestyle table-vaulting parkour performed autonomously
• World's first aerial flip with maximum height exceeding three metres
• World's first 7.5-rotation Airflare grand spin
• Maximum cluster movement speed of 4 m/s across a group of robots moving together

These are not simple motions. Parkour and kung fu require dynamic balance, rapid correction of mid-motion errors, and the ability to generate force through a kinematic chain in the right sequence at the right time. Doing any of this autonomously — without a human operator in the loop — requires a robotics stack that would have been considered advanced laboratory research two years ago.

Within minutes of the broadcast ending, Unitree robots sold out on JD.com. MagicLab and Noetix robots also sold out the same evening.

The reinforcement learning pipeline behind it

None of this happened because someone manually programmed each movement. The motions were learned, not coded.

The standard approach for training robots to perform dynamic motions has converged around reinforcement learning in simulation. You build a physics simulation of the robot body, define a reward function that gives points for completing the motion well and penalties for falling or using too much energy, and run the robot through millions of simulated attempts until it learns a policy that scores well. Then you transfer that learned policy to the real robot.

The transfer from simulation to reality — called sim-to-real transfer — is where many robotics companies have struggled. Simulations are not perfect. They do not perfectly model friction, motor backlash, and cable dynamics. When you take a policy trained in simulation and put it on a real robot, it often fails because the real world does not match the simulation it learned from.

Unitree has clearly made significant progress on this problem. The fact that complex learned motions are executing reliably on a stage in front of a billion viewers suggests their sim-to-real pipeline is working. That is a harder technical result than the performance itself.

What the commercialisation numbers mean

After the gala, Unitree announced plans to ship between 10,000 and 20,000 humanoid robots in 2026. They shipped approximately 5,500 in 2025. Morgan Stanley, covering the announcement, raised their forecast for total Chinese humanoid robot sales in 2026 to 28,000 units — up 133% year-over-year.

The market projections beyond 2026 are large. 262,000 units by 2030. 2.6 million units by 2035. These numbers assume continued cost reduction, continued capability improvement, and market development in manufacturing and logistics settings where humanoid robots have obvious applications.

For context on where this puts China relative to other robotics companies: Tesla's Optimus is still in limited trials at Tesla factories. Boston Dynamics' Atlas is not commercially available in the same volume. Agility Robotics is deploying Digit in Amazon warehouses, but not at the numbers Unitree is projecting.

The Chinese government has explicitly prioritised humanoid robots as a strategic technology. Subsidies, talent pipelines, and supply chain integration are all being coordinated at a national level in a way that has no equivalent in the United States or Europe.

What this means for developers working with robotics

If you are a software developer and robotics is anywhere in your professional horizon — as an area of interest, a potential employer, or a technology your company might eventually buy — the gala performance is relevant to your timeline.

The gap between what a developer with a robotics background could build two years ago and what Unitree is shipping now represents several years of progress happening in one. The reinforcement learning techniques behind the performance are published research — papers from ETH Zurich, Berkeley, and Unitree's own engineering team. The hardware is becoming available commercially.

The practical implication is that robotics development is moving from research institution territory into territory where a small team with the right skills can build something real and deployable. The tooling has matured: simulation environments like Isaac Gym and MuJoCo are more capable, model architectures like transformer-based policies are increasingly standard, and the hardware costs are dropping.

For developers who have been watching robotics from a distance, waiting for it to become accessible enough to take seriously: the Unitree gala performance is probably your signal that the wait is ending.

Why the West is genuinely behind

The more uncomfortable point is what the gala reveals about the gap between Chinese and Western robotics development.

The United States has significant robotics research capacity — Stanford, MIT, Carnegie Mellon, Boston Dynamics, and many startups. But research capacity is different from commercialisation speed. China's combination of hardware manufacturing infrastructure, government strategic prioritisation, competitive corporate investment, and willingness to deploy at scale in working environments has produced a different kind of progress.

The Spring Festival Gala is not just a technology demonstration. It is a cultural statement about normalisation. Showing autonomous humanoid robots performing kung fu to one billion people during the most-watched national broadcast of the year is a deliberate act of making humanoid robots feel like a normal part of the future.

The technology itself is real. The performance last February was not CGI or remote control. The robots did what was claimed. And the companies making them are planning to ship tens of thousands of units this year.

The West has strong AI research, competitive models, and significant robotics talent. What it does not have is the same combination of speed, scale, and national coordination that produced what happened at the Spring Festival Gala. Whether that changes in the next two years will determine a great deal about where the robotics industry lands.