Last month, dozens of Unitree G1 humanoid robots performed synchronized kung fu at China's Spring Festival Gala. They flipped, sprinted, handled nunchucks, launched off trampolines, and executed somersaults reaching three metres in height. The performance was broadcast to what became the most-watched live audience in 13 years — over 23 billion views across platforms. Within two hours, robot orders on JD.com surged 150%. Search volumes jumped 300%. Customer inquiries spiked 460%.

Unitree's CEO now projects 10,000 to 20,000 humanoid shipments in 2026, up from about 5,500 last year. Industry-wide, Chinese humanoid robot shipments are forecast to hit 62,500 units this year — a 270% jump. Clips of the robots went viral worldwide, and every tech outlet ran the same breathless headline: humanoid robots have arrived.

Have they?

I've spent 25 years manufacturing actuators — the component that makes a robot move, lift, push, and grip. I've watched this industry long enough to know the difference between a demo and a product. And what I saw in that viral kung fu performance wasn't a robot ready for your home. It was a robot performing on a clean, flat, obstacle-free stage under perfectly controlled conditions — doing something that has absolutely zero commercial application unless you're planning to start a robot boxing league.

The question nobody is asking is the only question that matters: how do you get from that to a robot that can actually work in your house without breaking your grandmother's hip?

The Clean Floor Problem

Every humanoid robot demo you've ever seen shares one thing in common: the floor is clean.

Watch any viral robot video carefully. The surface is flat, uniform, and clear of obstacles. The lighting is controlled. There are no pets. No toddlers. No shoes left in the hallway, no charging cables running across the kitchen, no wet tile in the bathroom. The robot knows exactly where it is because the environment has been engineered for the robot, not the other way around.

Now think about what "working in a home" actually means. A home is the most unstructured environment on the planet. Furniture gets moved. Kids leave toys on the floor. Dogs run underfoot. Rugs bunch up. Surfaces change from hardwood to carpet to tile within a few steps. Lighting shifts constantly. Doorways are narrow. Stairs are inconsistent. The kitchen floor might be wet. The bathroom might be slippery. A toddler might run directly into the robot's path while it's carrying a pot of hot coffee.

A robot that can do a backflip on a gymnasium floor and a robot that can safely carry a glass of water from the kitchen to the living room without stepping on a dog, tripping on a cable, or dropping the glass when a child screams — these are two entirely different machines. They share almost nothing in common from an engineering standpoint except the general shape of the body.

The backflip requires explosive power, precise timing, and a pre-calculated trajectory. It's a ballistic movement — the robot commits to a path and follows it. Domestic work requires the opposite: constant real-time adaptation, force-controlled interaction with fragile objects, and the ability to stop, redirect, or yield at any moment when something unexpected enters the robot's path. The actuator demands are fundamentally different. The sensor requirements are fundamentally different. The AI requirements are fundamentally different.

One is a trick. The other is a product. The industry is celebrating the trick.

The Skull-Fracturing Elephant in the Room

In November 2025, the former head of product safety at Figure AI — one of the most highly funded humanoid robotics companies in the world, valued at $39 billion — filed a whistleblower lawsuit alleging he was fired for warning executives that their robots were powerful enough to fracture a human skull.

According to the lawsuit, the company's Figure 02 robot generated impact forces during testing that were measured at twenty times the threshold of human pain, and more than double the force needed to break an adult skull. In one incident, a malfunctioning robot struck a steel refrigerator door hard enough to carve a quarter-inch-deep gash into the metal — while an employee was standing nearby.

The whistleblower, Robert Gruendel — a veteran safety engineer with experience at Amazon and BMW — claims his safety roadmap was gutted after it was used to help secure investor funding. His concerns were reportedly treated as obstacles rather than obligations. He was terminated days after filing his most documented safety complaints. Figure disputes the claims and says he was let go for poor performance.

Regardless of how the lawsuit plays out, the underlying physics are not in dispute. Humanoid robots are powerful machines. Their actuators are designed to generate the torque needed for locomotion, lifting, and manipulation. That same torque, in an uncontrolled interaction with a human body, is capable of inflicting severe injury. This is not a software problem. This is a physics problem. And physics doesn't get patched in a firmware update.

Figure's CEO, Brett Adcock, said something remarkably candid in a public statement: "I would not let my robot roam free for hours and weeks, right now, with my young kids." The CEO of a $39 billion humanoid robotics company will not let his own robot near his children. That should tell you everything about where this technology actually stands.

The Liability Chasm Nobody Wants to Talk About

Here is the billion-dollar question that I believe even the manufacturers themselves don't have an answer to: who is liable when a humanoid robot injures someone in their home?

Right now, there are no specific federal regulations governing humanoid robot safety in domestic settings. None. The first humanoid-specific safety standard — ISO 25785 — is still in development. Existing industrial robot standards like ISO/TS 15066, which defines force limits for collaborative robots, were written for factories where workers are trained, environments are controlled, and emergency stops are bolted to the wall. None of these standards contemplate a 70-kilogram bipedal machine walking through a living room occupied by a four-year-old and a Labrador.

When — not if — a humanoid robot injures someone in a home, the legal question becomes a nightmare. Is it the manufacturer's fault for a design defect? The software developer's fault for a bad AI inference? The actuator supplier's fault for a component that failed under an edge-case load? The homeowner's fault for leaving a toy on the floor that caused the robot to fall? What if the robot was running software from one company, actuators from another, sensors from a third, and an AI model from a fourth?

Product liability law as it exists today was not designed for autonomous machines that learn, adapt, and make real-time physical decisions. The legal frameworks that will eventually govern this space simply do not exist yet. And every manufacturer racing to put a humanoid into homes knows this. Their legal teams are, I suspect, having some very uncomfortable conversations right now about the liability exposure of selling a skull-fracturing-capable machine to consumers with no regulatory framework, no safety certification, and no legal precedent to hide behind.

This isn't just a barrier to market entry. It is potentially the single largest barrier to the entire humanoid home robot industry. The first major injury lawsuit — and it will come — will define how this market develops for a generation.

What "Crossing the Chasm" Actually Requires

The technology gap between a kung fu demo and a domestic utility robot is not incremental. It's categorical. Here's what actually needs to happen before a humanoid robot can safely do useful work in a home:

Force-limited actuators that won't kill you. Industrial humanoids are built with rigid, high-torque actuators designed for maximum performance. These are the actuators that carve gashes in refrigerator doors. Domestic robots need compliant actuators — actuators that can yield, absorb impact, and limit force output to levels that won't injure a human on contact. This is a fundamentally different actuator design philosophy. Some companies, like 1X with their NEO robot, are already designing lighter, softer robots for this reason. At 30 kilograms, a robot is far less dangerous than at 70 kilograms. But lighter robots also carry less, reach less, and do less. The engineering trade-offs are brutal.

Sensor systems that actually work in clutter. A robot on a stage has a clear field of vision. A robot in a home is surrounded by visual noise — furniture, clothing, reflective surfaces, transparent glass, moving shadows, pets that look like toys, and toys that look like obstacles. Current LIDAR and vision systems perform well in structured environments. In cluttered domestic spaces, they degrade significantly. The robot needs to distinguish between a child's arm and a broom handle in real time, in partial lighting, from an angle that was never in the training data. Getting this wrong isn't a minor error. It's a lawsuit.

Navigation that handles chaos. Autonomous navigation in homes isn't just pathfinding. It's predicting where a toddler is about to run. It's recognising that the cat sitting still on the kitchen floor might suddenly bolt. It's understanding that a pile of laundry isn't a wall and a glass coffee table isn't empty space. This is an order of magnitude harder than navigating a warehouse with painted lines on the floor. No humanoid robot has demonstrated reliable autonomous navigation in a genuinely cluttered domestic environment.

Safety certification that doesn't exist yet. Before a car goes on the road, it must pass crash testing, emission standards, and regulatory approval. Before a humanoid robot enters your home, it must pass... nothing. There is currently no domestic robot safety certification in any major jurisdiction. Until one exists — and until manufacturers are required to meet it — every robot sold for home use is essentially an uncertified prototype being operated by untrained consumers in uncontrolled environments. The insurance industry is already scrambling to figure out how to price this risk.

A legal framework for when things go wrong. Product liability, negligence, design defect, failure to warn — the legal concepts exist, but their application to autonomous domestic robots is entirely untested. Who bears liability for an AI decision? Can a manufacturer disclaim responsibility for an edge case the model never trained on? Is a homeowner contributorily negligent for not robot-proofing their home? These are questions for courts, legislatures, and regulators who have not yet begun to seriously address them.

The 3 D's: When the Real Market Opens

Twenty-five years ago, I coined a simple framework for evaluating whether any automation product has a right to exist commercially: Dirty, Dull, or Dangerous. If a robot doesn't do at least one of these things, it doesn't have a market — it has a fan base.

Everything we're seeing right now — the kung fu, the backflips, the synchronized dances — is entertainment. It's Phase 1. Early adopters buy humanoid robots for the cool factor, the same way people bought the first drones. The robots can walk, they can gesture, they can follow basic commands. But they can't do any of the 3 D's. They can't scrub your bathroom floor. They can't mow your lawn in the rain. They can't change your elderly parent's bedsheets. They can't do the jobs that would actually justify a $13,000 price tag to a mainstream consumer.

Phase 2 — utility — is when humanoid robots can reliably do at least one task that's Dirty, Dull, or Dangerous, in an unstructured environment, without injuring anyone or breaking anything, with enough consistency that the cost of the robot is justified by the labour it replaces. That's the real market. That's when unit volumes go from 62,500 to 6.2 million. And we are nowhere near that yet.

The companies that will cross this chasm are the ones that take the unsexy problems seriously: compliant actuator design, robust domestic sensor fusion, force-limited manipulation, safety certification, and liability engineering. Not the ones doing backflips for social media.

The Actuator Problem at the Heart of Everything

Every challenge I've described — force limitation, compliance, safety, precision in unstructured environments — comes back to the actuator. The actuator is the component that makes force. It's the component that can crush a skull or gently hand a child a glass of milk. The difference between those two outcomes is not the AI model. It's the actuator design, the control architecture, and the mechanical safety systems built into the hardware.

At FIRGELLI, we've spent 25 years engineering actuators with built-in end stops, force limits, and fail-safe mechanisms — because we learned decades ago that software fails, power cuts out, and edge cases happen. The question for the humanoid robotics industry isn't whether the AI is good enough. The question is whether the hardware is safe enough. And right now, for domestic deployment, the answer is no.

Unitree's kung fu robots are a remarkable engineering achievement. I mean that sincerely. The coordination, the speed, the balance — it's impressive. But impressive and safe are not the same word. Impressive and useful are not the same word. And impressive and commercially viable in a home full of vulnerable humans are very, very far apart.

The billion-dollar question isn't whether humanoid robots can do a backflip. It's whether they can do the dishes without landing someone in the emergency room. Until the industry can answer that, the kung fu is just a show.

And shows don't cross chasms. Engineering does.