November 4, Wuzhen. At the Internet Entrepreneurs Forum of the World Internet Conference, Wang Xingxing, founder of Unitree Robotics, took the stage. He spoke about what robots can do now, what they still lack, and which step has not yet been crossed.
The company has built robots for ten years, starting with quadruped machines and later humanoids. In this talk he discussed how much of a job a machine can do on a shop floor, and whether it can still do it somewhere unfamiliar.
One Hundred Percent at a Fixed Station
He began with the good news. In a single process, at a fixed station, a robot's task success rate can approach 100 percent if it is trained well. The weight of that sentence: factories need stability, and one failure can stop a line.
Then the bad news: efficiency still has to improve. Getting a motion right does not mean doing it fast, and doing it fast does not mean eight hours without error. A factory measures cycle time, not the one success in a demo video.
So he made the goal concrete: he hopes that someday, in any setting and any place, you can deploy a robot and it will work. Not a program written for one station, but something you can move and use.
A side note: why humanoid? Because factories, warehouses, stairs and door handles are designed around human height and human hands. Building a machine that fits existing environments is cheaper than rebuilding the environment. That is industry common sense, not his wording.
Eighty Percent in an Unfamiliar Room
This is the phrase he used last year: the "ChatGPT moment" for embodied intelligence. The standard: in about 80 percent of unfamiliar settings, a robot can complete about 80 percent of tasks through voice or text commands. Two 80s, one for settings and one for completion, neither negotiable.
One detail is easy to miss: he said voice or text commands, not pre-written programs. That raises the bar, because the same motion has to be summoned by a sentence, which means the model must first understand and then break the task into steps.
"Unfamiliar setting" deserves explanation. It does not mean another factory building; it means a room never visited, never calibrated, with a different table height and different light. The robot goes in and has to look, judge and walk over by itself.
Note that this is a goal, not a scorecard. His account: the world's leading multimodal models are already quite close, with strong planning in physical space, and relatively high success rates for grasping objects and simple carrying.
"Quite close" is the key phrase. It places the claim just before the threshold: not across it, not far from it. People who build technology use this phrasing because it is closest to the facts.
Short by the Last Few Millimetres
So where is the gap? He gave a short answer: the last few millimetres. It is not that the model is not smart enough; between the model's output and a real robot's motion there is a distance that has not been aligned.
His own explanation: AI still does not handle contact with the physical world well enough, and the alignment between its input and output and the real world is slightly off. In a workshop that means: you tell it to grasp, and it misses by two millimetres.
Two millimetres is nothing on a screen; on an assembly line it means the part will not go in. A robot can see, but not precisely enough; it can touch, but not feel finely enough. Vision has come a long way; touch is still at the door.
The Difficulty of a Needle
His examples were everyday ones: taking a phone apart, putting one together, threading a needle. They share a trait — small spaces, small parts, small tolerances — and they rely not on force but on feel. What decides the attempt is the sensation at the fingertip in the final moment.
When a person threads a needle, the eye cannot really see whether the thread has entered the eye; the fingertip's feedback does the work. What the robot lacks is that feedback, and the speed to send it back into the decision loop in time.
He also noted that basic tactile capability and the real-time transmission of control signals still lag the real world. But he added: it is just a little bit. That is a judgment, not a guarantee.
His estimate: in the next one to three years, if that tiny deviation can be solved, embodied intelligence will have no major obstacle in theory. Note "in theory" — mass production is still some distance away.
Most of the Data Is Still Unused
Besides touch, he raised another problem: low data utilisation. At Unitree and across the field, large amounts of collected data are not fully used; they sit there. The data is not useless; what is missing is a way to string it together.
A few months ago, the company began experimenting with self-evolution for physical models. The model writes its own code, trains in simulation or the real world, deploys to a robot for testing, and then a human or a robot scores the result, which is fed back to revise the program.
The loop sounds much like what large-model companies already run internally. In a sense, he said, leading model companies have achieved self-evolution, with humans only scoring a small part; in robotics it is still rare.
Why rare? Because robots have one more gate: a motion that works in simulation may be a different thing on the real machine. The gap between simulation and reality is exactly those millimetres.
The Timetable He Gave
Last, time. He believes there is a high probability that next year will be the inflection point for the robotics industry or for robot models. That is a judgment, not a promise; his own words were "high probability."
If the last few millimetres are smoothed out in the next year or two, the change will not stay in labs. Factory shift schedules, warehouse sorting and mall guidance will feel it first, because those environments are relatively fixed.
The reverse also holds. If the millimetres are not smoothed out, robots will stay at fixed stations, performing motions they have been trained on again and again. It looks busy, but the arithmetic may not work.
Three years is not long in robotics. Software iterates by the week, hardware by the year; one leg fast and one leg slow makes you walk in circles, so the two have to move together.
He also gave a figure: in recent years, nearly half the company's robots, on average, have been exported each year, and many overseas firms buy them for secondary development in construction and commercial service. Demand is real; it is just not yet general.
So the line worth remembering from this talk is not the 80 percent goal but "the last few millimetres." It shrinks a very large subject to a measurable question: can a machine pass a thread through the eye of a needle.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.