VERTEBRATE.ai

Unitree, the Hangzhou-based robotics maker known for its backflipping humanoids and bargain-priced robot dogs, is now working on something less photogenic but potentially more consequential: an AI system that would let its machines get better at their jobs without an engineer standing by to patch the code. The company says the system is designed to let humanoid robots learn from their own performance and refine their movements over time, rather than waiting for each new skill to be hand-programmed and pushed out in an update.

The disclosure came alongside remarks from founder Wang Xingxuan, who told reporters he expects the humanoid robot industry to reach its own "ChatGPT moment" — a breakthrough so sudden and legible that it resets public expectations — by 2027. It was Wang's framing, not a marketing slogan, and he paired it with an unusually candid caveat: even a leap that dramatic won't by itself solve the harder problem of turning a capable robot into something companies and consumers actually want to buy and deploy at scale.

That combination — technical bravado married to commercial humility — is the real story here. Unitree has spent the past two years building a reputation as the industry's fastest mover, undercutting rivals on price while matching or beating them on acrobatics. A self-improving AI layer would push that advantage further, because the costliest, slowest part of deploying robots in warehouses, factories or homes isn't the hardware anymore; it's the endless cycle of reprogramming machines for every new task, every new environment, every new edge case. A robot that can watch itself fail and adjust, the way a toddler learns to stack blocks, would cut that cycle dramatically.

It would also change the economics of the business. Robotics companies today sell largely on specifications — torque, battery life, range of motion. A system that lets robots accumulate competence after they've left the factory shifts the value proposition toward something closer to what software companies sell: a product that improves on its own, long after purchase, the longer it's used. That's the pitch behind every "ChatGPT moment" analogy in robotics right now, and Wang's version is notable mainly for how unsentimental it is about the distance between a lab demo and a loading dock.

For an industry still haunted by over-promised, under-delivered humanoid robots — Boston Dynamics parkour reels that never became products, Tesla's Optimus timelines that keep slipping — Wang's acknowledgment that capability and commercialization are separate problems reads less like hedging and more like realism. If Unitree is right that 2027 brings a genuine capability jump, the harder test will be the one that follows: whether factories, hospitals and households are actually ready to let a self-teaching machine loose among them, and whether anyone's figured out how to price, insure, and regulate a robot that keeps changing after it's sold.