Strictly speaking, the United States did not ban Unitree robots at the border. On 28 July the Federal Communications Commission placed all foreign-made advanced robot devices under control, so new devices in principle cannot get FCC authorisation without conditional Pentagon approval, while already-certified models keep selling. For Unitree, racing toward a STAR-market listing, it is still a suddenly raised bar.
Not a border ban
Unitree secured CSRC approval for its listing on 2 July, but under the FCC rule its new US-bound products face fresh access questions. It flagged the US market risk in its updated prospectus on 31 July. The US was 18.39, 19.54 and 13.30 per cent of revenue in 2023, 2024 and the first nine months of 2025, a meaningful share. American universities and robot labs do more than order: they write algorithms, publish papers, generate training data and pull more developers onto the same platform. If new models cannot enter that research market, Unitree loses a developer network that shapes future standards, not just a batch of orders.
The platform trap
Unitree’s success rests on ‘high performance, low price, standardised platform’. It never chased one giant end scenario; it sold quadruped and humanoid bodies to universities, labs, tech firms and individual developers who built the applications. In the first three quarters of 2025, research and education were 73.6 per cent of humanoid revenue, commercial consumer 17.39 per cent and industrial only 9.01 per cent, with real smart-manufacturing and inspection revenue about 2.6 per cent of the humanoid total. That is a viable business, selling a robotics development platform, but it is also the obstacle to the next stage.
From demo to work
A platform company sells robots; an embodied-AI company sells the robot’s completed work and owns the result on site. The maths of long tasks is brutal: at 99 per cent success per step, twenty steps in a row finish only about 82 per cent of the time, and at 95 per cent just 36 per cent. Unitree has proven motion control, but those skills do not directly translate into stable completion of complex jobs. It is building data capability, with a G1-D pipeline for concurrent collection and the HIW-500 set, 500-plus hours and 23,000 episodes across 12 real homes, yet the gap from single-step imitation to long-horizon autonomy remains the industry’s hard wall.
Read the original report (Sohu IT)
Translated and adapted from Sohu IT (it.sohu.com).