One year ago, Unitree’s robots wobbled through a Yangko folk dance on the world’s most-watched TV stage.

Sixteen humanoids.

Spinning handkerchiefs.

A charming, slightly shaky performance directed by Zhang Yimou.

Cute, but fragile.

(Video source: CGTN YouTube Channel)

Fast forward to February 17, 2026. Same stage. Same company.

Twenty-four humanoids executing 3-meter aerial flips.

Single-leg backflips — three in a row.

An airflare spin of seven-and-a-half rotations.

Cluster repositioning at 4 meters per second.

Wielding swords and nunchaku alongside child martial artists from Henan Tagou, one of China’s most prestigious kung fu schools.

The performance, titled “WuBOT” (武BOT), aired to an estimated 679 million viewers.

(Video source: Unitree YouTube Channel)

Weibo erupted. One comment cut through: “Last year it could only twirl a handkerchief.”

Well, the skeptics arrived immediately:

“Ha, Pre-programmed. Remote-controlled. Performance art, not real AI.”

Here’s my take: the skeptics are only half-right. They are arguing about the wrong problem.

The hardware is solved (or very close). And at a price that changes everything.

For three decades, bipedal locomotion was the hardest unsolved problem in robotics.

Not software. Hardware.

Building a physical body that could walk, absorb ground impact, maintain balance under unpredictable load, and survive real-world use — that problem consumed billions of dollars and entire research careers.

Boston Dynamics spent 30+ years on it. Honda’s ASIMO spent 18 years learning to walk stairs slowly and pour a drink. It never shipped commercially. (Eighteen years! My entire adolescence.)

The physics are brutally unforgiving. A 50kg robot falling on a factory floor is a $90,000 pile of destroyed components.

That’s the problem Unitree solved. Cheaply.

Their G1 humanoid starts at $13,500. Not $1.3 million. Not $130,000. “Thirteen thousand, five hundred dollars” — the price of a used car.

How?

Full vertical integration.

Motors, reducers, actuators — everything designed in-house. No external supplier taking margin. No off-the-shelf components built for someone else’s specs.

Wang Xingxing, Unitree’s founder, explained the logic plainly: reduce wires, reduce chips, reduce screws. These seem basic, but they compound. Although it sounds counterintuitive, designing motors and reducers in-house delivers both better performance and lower cost than buying off-the-shelf.

This is the same playbook that made BYD dangerous in EVs: own the supply chain, control the cost structure, price the competition into a corner. According to the industry research firm GGII, Unitree sold 23,700 quadruped robots in 2024 alone — capturing nearly 70% of the global market.

To see how Unitree’s pricing compares to everyone else in the humanoid race:

The skeptics aren’t wrong about pre-programming. The performance ran on controlled terrain, known choreography, and scripted sequences. Even the one planned “fall” — a robot stumbling and being helped offstage — was deliberate. Wang (short for Wang Xing Xing) joked to CCTV afterward: “Actually, it was just for fun.”

Three systems made the performance possible. Each one matters more for what comes after the stage than what happened on it.

The hard problem in cluster robotics: keep 24 robots precisely positioned while they are flipping and sprinting — without losing track of where they are.

Traditional positioning algorithms drift. A robot sticks a landing 2 centimeters off. Fine alone. In a cluster moving at 4 m/s, that error compounds into collision.

Unitree fused two data streams: proprioceptive data (the robot’s internal sense of its own joint positions and forces) and 3D LiDAR (laser-based environmental scanning). An AI model processes both hundreds of times per second, maintaining precise positioning through dynamic movement. No external motion capture systems. No GPS. Just onboard intelligence.

Think of it like a gymnast using both muscle memory and spatial awareness simultaneously. Remove either input, and they fall.

Wang told CCTV: “This year, the robots completed interspersed formations and martial arts actions while running fast. This high-dynamic and high-coordinated swarm control technology is making its global debut.”

Each robot executes individual commands while staying synchronized with the group. Motion synchronization error was strictly controlled to ultra-low latency. Twenty-four autonomous bodies, one choreographic mind.

The three consecutive single-leg backflips drew the most attention. The team built a custom launching platform — similar to assistive devices humans use for extreme maneuvers — letting robots leap 2–3 meters in the air and land clean.

Behind that 30-second sequence: hundreds of millions of training iterations in simulation before a single rep on physical hardware. Wang told CCTV that this maneuver places extremely high demands on balance control, dynamic response, and landing stability. Then, quietly: “Up close, it can almost jump as high as the ceiling.”

If you are still not quite sure what I am talking about, watch the Unitree video again at the beginning.

Here’s the part the skeptics miss. All three systems transfer directly beyond the stage:

Replace “kung fu choreography” with “warehouse logistics” and you would have the same engineering problem!

It was almost midnight when the 36氪 (36kr) reporter caught up with Wang Xingxing backstage at the Xingguang Film and Television Park in Beijing’s Daxing district.

He’d just delivered what became China’s most-talked-about tech moment of the year. The Gala was still airing. He was still wired. And when the conversation turned to technical details, the usually introverted founder kept demonstrating the robot’s fighting moves with his own body — voice rising, arms swinging — as he spoke to reporters past midnight.

“After the performance, a big stone in my heart finally dropped,” he told 36kr. “Over the past one or two months, I personally felt a lot of pressure. We wanted to present something far better than last year.”

That pressure has a name: the brains haven’t kept up.

Wang himself said it in a keynote speech: large models for embodied intelligence are still far behind hardware. Robots still look more like expensive toys than labor replacements.

The Gala — spectacular as it was — exists in a comfort zone. Controlled environments. Known terrain. Scripted sequences. Tell one of those robots to clean your living room. Your living room is chaos. Objects move daily. Commands are ambiguous. The line between trash and your dog’s favorite toy isn’t written anywhere.

Goldman Sachs analysts reviewing the broader industry agree: impressive hardware, nowhere near mature enough for serious industrial deployment at scale.

Wang knows this. He said so publicly. Which is exactly why the midnight scene matters: a founder who just delivered China’s biggest tech moment, still shadow-boxing robot moves at midnight, already thinking about what the brains need to catch up on.

Here’s what I think the Gala was actually about — and why the pre-programming argument misses the point entirely.

Wang told 36kr something that Western coverage mostly ignored:

“These technologies are very practical and conducive to large-scale cluster work of robots in the future.”

He wasn’t talking about performance art. He was talking about factories. The stage was a proving ground — a live stress test for cluster control and rapid repositioning systems that transfer directly to industrial deployment. (That’s ¥100 million — about USD $14 million — to put robots on TV. Expensive product demo. But when 679 million people watch your technology work, the marketing ROI is hard to argue with.)

But there’s a deeper strategic move happening.

This is just my thesis, but: Unitree is building the Android of robotics.

Android didn’t win because it was more sophisticated than iPhone. It won because it was open, cheap, and everywhere — which attracted every developer on earth to build on top of it. Android now runs on 72% of smartphones globally.

Unitree’s G1 is following the same playbook:

The critical development that makes the Android thesis viable: robot AI models are increasingly designed to transfer across hardware.

Google DeepMind’s Gemini Robotics — launched March 2025 and iterated through the year — runs as a vision-language-action (VLA) model explicitly evaluated across multiple robotic platforms, from bi-arm ALOHA setups to Apptronik’s Apollo humanoid. It can be fine-tuned with as few as 50 demonstrations. Stanford’s OpenVLA was trained on data from 22 different robot embodiments. Hugging Face’s SmolVLA runs on a laptop. At CES 2026, Google DeepMind and Boston Dynamics announced a partnership to integrate Gemini Robotics into the Atlas humanoid.

If that trend continues — and the scaling research suggests it will — the company that puts cheap, capable, open hardware into the most researchers’ hands wins the data flywheel:

Here are some numbers that back this thesis:

This isn’t a startup burning VC money chasing demos. It’s a profitable business scaling into an Android position.

Intellectual honesty requires acknowledging the other side.

Figure AI raised over $1.9 billion. Their valuation hit $39 billion after a September 2025 Series C. Their robots haven’t performed kung fu on TV. However, they spent 11 months inside a BMW factory in Spartanburg, South Carolina.

Two robots. 10-hour shifts, Monday through Friday. 1,250+ hours of runtime. 90,000 sheet metal parts loaded. A contribution to 30,000 BMW X3 vehicles.

The task was unglamorous: pick a sheet-metal part from a bin, place it on a welding fixture within a 5-millimeter tolerance. Complete the cycle in 84 seconds. Repeat. Every day. For months.

The robots came back scratched and battle-worn. Figure showed the photos deliberately — the grime was proof, not embarrassment.

And 1,250 hours of factory runtime bought something no demo generates: the forearm became their top hardware failure point.

Tight packaging, thermal constraints, constant motion stress. For Figure 03, they completely redesigned the wrist electronics. Each wrist motor now communicates directly with the main computer. That design fix would never have surfaced on a Gala stage.

More importantly: every hour generated proprietary training data that no competitor can replicate without doing the same thing.

Tesla runs the same playbook — deploying Optimus into their own factories specifically to harvest training data.

The real product isn’t the hardware.

This is the Apple bet: own the hardware, the brain, and the training data. Closed loop. Potentially unbeatable moat if integration advantage compounds.

So which bet wins? Here’s how they stack up:

The case for Apple: Robot AI isn’t like a smartphone app. Training data collected on Figure’s actuator dynamics doesn’t transfer perfectly to Unitree’s different joint architecture. Integration matters in ways pure software modularity doesn’t fully capture. Figure dropped its OpenAI partnership to build its own in-house model (Helix), specifically because they believe the hardware-software marriage is the winning formula.

The case for Android: Slow, expensive, geographically constrained. Getting two robots into BMW took months of setup. Figure hasn’t disclosed whether their robots met the 99% accuracy and zero-intervention-per-shift targets. Scaling to 100,000 units — which CEO Brett Adcock has targeted within four years — requires manufacturing scale Figure is only beginning to build. Meanwhile, Unitree ships 10,000–20,000 units in 2026 to researchers worldwide generating data on their own problems.

Both bets can win simultaneously, in different markets.

Android has 72% of global smartphone market share.

Apple captures most of the profit.

Both survived. Both are worth trillions.

But I lean toward the Android bet having larger structural impact. Here’s why.

The “ChatGPT moment” for robotics — the point where robots generalize across tasks in unstructured environments — requires training data at a scale no single company can generate alone. Google DeepMind trained Gemini Robotics across multiple robot embodiments specifically because more diverse hardware generates more generalizable models. The company that distributes the most hardware creates the largest, most diverse training dataset.

That company, right now, is Unitree.

Wang himself acknowledged this dynamic at the World Robot Conference. He said he was skeptical of current VLA model architectures, calling them “relatively simplistic.” But he’s hedging — collaborating with third-party companies on embodied AI while maintaining an open attitude. In robotics, he admitted, there’s an element of luck in model breakthroughs. You don’t bet on one architecture. You bet on being the platform others build on.

One year ago, 16 Unitree robots twirled handkerchiefs.

This year, 24 of them executed autonomous kung fu in front of 679 million people, while their CEO stood backstage past midnight, shadow-boxing the same moves, already thinking about what the brains need to catch up on.

The hard part is already over.

The interesting part starts now.

The hardware race is largely settled. Competition shifts to foundation models for physical AI — models that generalize across real-world environments. Wang’s own “ChatGPT moment” prediction: 2–3 years at the earliest, 3–5 years at the latest. Watch what Google DeepMind, Chinese AI labs, and startups like Physical Intelligence ship for robot reasoning this year. The brain gap Wang acknowledged publicly is where the next war gets fought.

2 → Unitree’s IPO Filing expected on Shanghai’s STAR Market.

When shipment volume, revenue split, and gross margins become public, we’ll know whether the Android bet works as a business — or whether Chinese price wars destroy margins before the market matures. Morgan Stanley’s forecast of 28,000 Chinese humanoid units in 2026 sets the demand baseline. Watch whether Unitree captures the lion’s share or gets undercut by AgiBot and newcomers.

Track real-world deployments. Unitree robots are reportedly in BYD and Geely EV factories. Figure completed its BMW pilot and is deploying Figure 03 next. The companies logging unglamorous factory hours are building the data moats that compound. Gala performances are marketing. 1,250 hours of sheet-metal loading is a moat.

Cheers.

Zero Address covers Chinese technology for English readers. I read the Mandarin tech docs so you don’t have to.

If this reframed the story for you — share it with one person who is still reading the headline version. And of course, subscribe to receive free articles like this every week.