Unitree's founder tested GPT-6 on robots ... the results might have been too good to report
Unitree founder Wang Xingxing stood at a roundtable at the China International Industry Fair in Shanghai on Monday and said something pretty significant: after testing GPT-6 on his robots, the results were "very, very good."
Now he's moved his embodied AI tipping point forecast from two to three years down to one to two. That's the time when physical AI is so good robots can go into unfamiliar environments and accomplish tasks without extensive retraining.
The caveat: he published no numbers. No task descriptions, no success rates, no baselines. A 90-second clip from The Paper is the primary source. What Wang offered was conviction, not data, but conviction from the person who sold more humanoid robots than anyone else on earth in 2025 is probably worth paying attention.
The key significance here is that there is a school of thought that general AI foundation models will get so good that general robot training will essentially soon be a solved problem: out of the box, no massive and expensive training program needed, just some fairly quick on-the -job training for specific tasks.
What he actually said
Humanoids Daily quotes Wang directly:
"We also put GPT-6 on all kinds of robots and tested them on all kinds of tasks, and the results were actually very, very good."
On the tipping point, he said: "It could even be one or two years." In August at the World Robot Conference, his fast scenario was two to three years and his slow scenario was five to ten, for a system capable of completing roughly 80% of requested tasks across 80% of unfamiliar environments. The Shanghai remarks reveal a more optimistic horizon, though he didn't restate the 80/80 threshold explicitly.
He also speculated that GPT-7 might arrive next year. "It could break through the field's remaining problems," he said.
The most revealing line: "The biggest mistake I made over the past decade or so was not believing in AI enough."
The actual GPT-6 robot evidence
Wang's tests aren't published, but GPT-6 Astra's robotics performance has been tested independently.
The picture is mixed.
RoboCurve ran a series of manipulation trials. On a simple block-in-bowl task, Astra succeeded on 19 of 20 attempts. On precision puzzle insertions, it managed 2 of 20. That spread reveals the core problem with general-purpose AI models in robotics: they work well on tasks that tolerate imprecision and fail on tasks that require it. A robot sorting objects into a bin is a different problem from a robot assembling a component.
One can absorb errors. The other cannot.
Previous research showed that combining Astra with a specialist robotics policy improved performance on harder two-arm tasks. That's the more likely deployment architecture: a general model for planning and language reasoning, a specialist policy for precise manipulation. NVIDIA's Long-WAM work showed a similar pattern, where giving a robot up to 38 seconds of video context dramatically improved task performance, but the gains were most pronounced on structured tasks, not open-ended ones.
Why Wang's timeline matters even without the data
Unitree shipped roughly 5,500 humanoid robots in 2025.
That's more than any other humanoid manufacturer on the planet, and it's not close. Unitree's G1 robots are deployed in entertainment venues, exhibitions, and commercial spaces across China and globally. When Wang Xingxing updates his deployment forecast, he's updating it from a position of operating scale that no one else in the industry has, at least right now.
That makes his GPT-6 comment structurally important, even without the metrics.
He is literally one of the best-positioned people in the world to know what general-purpose AI models actually do when you run them on real hardware doing real jobs. If those tests were producing results that contradicted his optimism, he would not be standing at a public industry event shortening his timeline.
The absence of evidence is not evidence of absence, sure, but the direction of Wang's update is a very, very big hint.
He also acknowledged the problems that still exist.
"Generalization and efficiency still need improvement for large-scale adoption," he said.
Robots that work in one environment need to be retrained or reconfigured for the next. The cost of that retransfer is the real bottleneck for industrial deployment, not the quality of any single demonstration. China's rental humanoid market exposed exactly this gap: robots that perform well in demo conditions require significant intervention in real operating environments.
The question the clip doesn't answer
GPT-6's robotics capabilities are real.
The independent tests confirm it.
What Wang's comment doesn't address is the gap between "impressive test results" and "deployable at scale." Those are different bars. Humanoid robots are already shipping at scale without systematic safety validation. Adding a more capable general AI model into that stack accelerates the performance curve while also potentially accelerating the failure modes.
Of course, the 80/80 threshold Wang uses is not a formal benchmark.
It's a conceptual description of when the average task in the average new environment succeeds without human intervention. Getting from 19/20 block-in-bowl to 80% of tasks across 80% of new environments is a big step.
Is it big enought?
Wang's 1 or 2 years is aggressive. But so is the pace of improvement of today's frontier models.
Sources
Written by
I tell stories with data and build startups. I've raised angel and VC funds, led software companies and startups, and chronicled the rise of the mobile economy. Currently I write for Forbes, host several podcasts, and am starting up HumanoidMegaHub. I was TUNE's mobile economist and Singular's VP of Insights, built the research team at VentureBeat, and managed teams creating software for partners like Intel and Disney. In past lives I built websites, e-learning platforms, and mobile apps, and, a long time ago, was named to Folio's top 100 of the media industry's "most innovative entrepreneurs and market shaker-uppers.”