Here's Why a Robot That Can Fight Kung Fu Still Can't Sort a Crate

Here's Actually the Difference Between a Robot That Dances and a Robot That Works

A kung fu routine and a dance are choreography - a fixed sequence of movements, rehearsed until the robot's control system can execute them reliably, the same way every time, in a controlled setting. Sorting a crate is not choreography.

It requires the robot to look at an object it hasn't seen positioned exactly that way before, decide how to grip it without dropping or crushing it, adjust in real time if the grip slips and repeat that judgment call differently for the next item.

The gap between those two tasks is the entire story here and it has a name: embodied AI, which just means artificial intelligence built into a physical body so it can perceive its surroundings, decide what to do and act - as opposed to a chatbot's AI, which only has to process text.

Where the "IQ Is Too Low" Actually Bites

China's robot hardware is genuinely excellent - the country produced roughly 95% of the 20,000 humanoid units shipped worldwide in 2025 and its manufacturers can build extremely capable physical machines. What those machines lack is generalised judgment: the capacity to handle a situation slightly different from what they were trained on.

At the Liuzhou facility, human trainers wearing headsets essentially puppet the robots through tasks - seeing through the robot's cameras, translating their own arm movements into the robot's motors - to generate training data for the AI models underneath.

Tang Wenbin's blunt assessment, "a lot of what we see is dancing disguised as working," captures exactly this: a robot that can be taught one precise choreographed sequence isn't the same as a robot that can handle the variability of an actual factory floor, which is why a novice trainer needs roughly 300 attempts to produce a single usable movement.

Why China Is Scaling Hardware Anyway

• And Where India Stands Here is the genuinely strategic bet: Beijing is applying the exact playbook that worked for EVs and solar panels - subsidise production early, build capacity before demand catches up and let fierce domestic competition drive down costs - to humanoids. That worked for EVs and solar because those were mature technologies with a known, solvable engineering path. Humanoids are different: the hardware is largely solved, but the software - general-purpose embodied intelligence - is not and no amount of subsidised factory capacity accelerates a machine-learning research problem the same way it accelerates battery manufacturing. India, by contrast, has no comparable humanoid manufacturing base and is not attempting to compete on hardware scale; the more relevant Indian angle is the AI research layer itself, where the actual bottleneck China is racing to solve currently sits and where compute access and algorithmic research - not factory capacity - determine progress. For an aspirant, the exam-relevant insight is precise: don't mistake manufacturing dominance for technological leadership when the constraining bottleneck is a different kind of capability entirely - here, hardware scale and software intelligence are simply not the same race.