You’ve seen the viral videos. A $100,000 humanoid robot executes a perfect martial arts form, or does a backflip, or delicately folds laundry. The crowd goes wild. The investors nod. And then… nothing. The robot goes back in the lab.
We are living through the greatest hype cycle of our generation, and it is built on a fundamental misunderstanding of what a robot is supposed to do. Everyone is obsessing over making machines that look like humans, completely ignoring the fact that a robot that can do a backflip is a science project. A robot that makes four million cups of coffee is a business.
Tang Mu knows this better than anyone. He spent a decade at Tencent building internet products, then moved to Xiaomi to build hardware like the Xiaoai smart speaker. Now, he’s the founder of XBOT, building robotic coffee shops. He didn’t build a humanoid. He built a machine that does one thing flawlessly: serve high-quality coffee 24/7.
His company has deployed over 1,000 robot baristas globally. They haven’t just made a few lattes for a trade show demo; they’ve produced over 4 million cups of coffee in the wild, facing real customers, real supply chain issues, and real rent pressures.
The chasm between embodied AI and commercialization isn’t a technology problem. It’s a product definition problem. The word “humanoid” is a trap. It tells you what the machine looks like, but it doesn’t tell you what problem it solves. A vacuum robot cleans floors. A coffee robot makes coffee. What does a humanoid robot do? Everything? Nothing?
The most dangerous thing a product can do is fall short of expectations. And a humanoid robot promises the entire sci-fi universe.
When you promise a sci-fi utopia, you deliver a parlor trick. The market doesn’t care about parlor tricks anymore. The market cares about ROI, operational consistency, and whether your machine can survive the brutal, unglamorous reality of 7×24 operations. Can your robot handle a clogged ice machine? Can it maintain output consistency across a thousand locations? If your PM is still obsessing over whether the robot can do a somersault, you are building a fake product.
And this brings us to the biggest lie in the AI hardware space today: the myth of data collection. Everyone thinks deploying robots is just a sneaky way to gather training data for the grand, unified AI model. That’s backwards.
Operational data isn’t training data. It’s the lifeblood of a product that actually has to pay its own rent.
The data generated from a real commercial deployment—customer orders, equipment failures, peak-time traffic, supply depletion—isn’t just raw material to feed into a neural network. It is the immediate feedback loop for product iteration, operational efficiency, and business survival. If you treat real-world data merely as training fodder, you ignore the gritty reality right in front of your face. You have to manage data as a product, one that dictates whether your physical business lives or dies.
This reality is fundamentally changing the role of the product manager. The days of the document-writing PM are dead. AI tools like Codex and Claude can spin up prototypes in hours. The traditional PM who just translated a boss’s idea into a PRD is obsolete. The new PM must be a full-stack builder. They need to integrate AI tools, understand supply chains, calculate ROI, and make hard calls on what actually gets shipped.
But don’t confuse AI output with product judgment. AI can generate a thousand variations of a feature in seconds. It can analyze operational data and surface insights. But AI can generate a prototype in seconds, but it still can’t tell you if anyone actually wants to buy it. The ultimate scarcity in the AI era isn’t compute power or model output. It’s human value judgment. Which problem is actually worth solving? Which edge case matters? Which business model is sustainable?
If you want to survive the embodied intelligence wave, you have to stop chasing the “smart home hub” or the “humanoid assistant.” Stop forcing new technology into scenarios where it doesn’t belong—nobody wants to yell at a smart toilet to flush. Technology is just a tool. A product must be explainable in one sentence, and once people use it, they shouldn’t be able to imagine going back.
Building a robot is a demonstration of capability. Making a robot that people actually use, day in and day out, is the creation of value. Stop building for the demo reel. Start building for the real world.
FAQ
Q: Isn't building general-purpose humanoid robots the ultimate goal, even if it takes longer?
A: It's a massive trap. 'Humanoid' is a marketing term, not a product category. If your robot doesn't have a specific, high-frequency task to execute, you are just building an expensive demo. The market pays for solved problems, not sci-fi aspirations.
Q: What should Product Managers focus on in the embodied AI era?
A: Stop being a PRD translator. AI can write the code and generate the prototypes now. PMs must become full-stack builders who understand supply chains, calculate ROI, manage 7x24 operational data, and make the final human value judgment on what is actually worth shipping.
Q: If I deploy robots in the real world, isn't the main point to collect training data for the AI model?
A: No, that's a dangerous backwards view. Operational data is not just training data. The most valuable real-world data must first serve product definition, operational iteration, and business survival. If you ignore the immediate operational ROI to chase future model training, your business will go bankrupt before the model ever finishes learning.