The demonstration is small and the implication is not. A general model that was never trained for robotics picked up a robot arm costing a few hundred dollars, held a brush, and drew the Golden Gate Bridge on paper. The first attempt was crooked. By the fourth, the outline and detail were legible.
The video was posted on X by Thijs Simonian, who works on the OpenAI robotics team. The argument that followed has not settled since.
The claim being tested
The strongest version of it, repeated across Chinese and American technical circles, is that the OpenAI of robotics will turn out to be OpenAI, rather than Physical Intelligence or Skild.
Until now the embodied brain thesis was hard to falsify. Models were evaluated on curated tasks, demonstrations were edited, and the volume of real robot data collected became a valuation argument in itself. A general model producing competent manipulation without any robotics fine-tuning sets a reference point that the sector has been missing.
What Chinese builders found when they tried it
Unitree was among the first to discuss Astra openly. Founder Wang Xingxing said his biggest past mistake was not believing in AI hard enough, and predicted that next year’s model generation could break the embodied problem open.
Without any robot-specific training, Unitree used Astra to let a machine identify a paint palette, change pigments and clean a brush. More complex tasks, such as tidying a room on instruction, may be within reach next year.
SenseTime ran its own evaluation and concluded that general models can now handle manipulation and navigation work previously delegated to specialised vision-language-action and world action models, showing zero-shot transfer, in-context learning and recovery from failure.
For leading body makers such as Unitree and AgiBot, that cuts both ways. Their advantages in mechanical performance, motion control and unit cost get flattened if the brain becomes a commodity. Their bodies, meanwhile, become the scarce asset, because intelligence still has to be delivered into the physical world at a reliable cost.
AgiBot co-founder Peng Zhihui has argued that intelligence has to land on a dependable body either way. At the company’s 29 October innovation day he went further, saying AgiBot can also be an AI company. Supporting 30 million levels of embodied tasks, its training cluster is now close to 10,000 accelerators.
The parts that do not work yet
Astra is not universal. Cloud inference carries latency that rules out tight control loops, and success rates on force-feedback tasks requiring high dynamic precision are close to zero.
That residual is the whole business case for the specialists. Fine assembly, contact-rich manipulation and anything where a millisecond matters still belong to models trained on the machine doing the work.
The uncomfortable conclusion for the standalone embodied brain startups is different. If a frontier lab can reach competent manipulation without a robotics programme, the moat they were raising money against was never a moat. It was a timing gap.
Editor’s note: This is an adapted translation of the original Sohu Tech report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1097481126_120924473.