At the IEEE's flagship robotics conference in Vienna, embodied AI (meaning robots that learn by physically acting on the world) was the dominant theme among more than 30 Chinese exhibitors, and the pattern the show revealed was structural: in immature AI hardware sectors, the…
Vienna's big robotics conference last week offered a clean view of a pattern shaping AI-hardware industries: when a field has no mature supply chain to specialize against, the companies that build every layer themselves usually outcompete the specialists. China's embodied-AI (具身智能) startups, the most visible current example, came to ICRA 2026 doing exactly that, but the pattern itself is portable to humanoid robotics, lab automation, and autonomous systems everywhere.
The IEEE's flagship robotics research conference ran June 1-5 in Vienna, and a reporting team from Leiphone's AI科技评论 trade desk walked more than 30 Chinese-company booths over five days. The walkthrough, published July 1, surfaced three threads that fit a single underlying story.
The first trend is the collapse of specialization into full-stack integration. Robotics has historically divided work: some companies build bodies, others build dexterous hands (灵巧手), others write the control software. At ICRA 2026, according to the Leiphone walkthrough, that boundary was visibly dissolving. Body makers including 它石智航 and 星动纪元 now also build hands, while hand and joint makers such as 舞肌科技, 灵心巧手, and 源升智能 now build whole bodies and the AI models that drive them. The reason is structural rather than strategic. The industry is still early, with no standardized division of labor, so companies plug their own capability gaps because no one else will.
The second trend is that data capture has moved to center stage. Leiphone's reporting describes 数据采集工厂, or data-capture factories (dedicated facilities where robots perform actions to generate training data), run by 千寻智能, 零次方, and 魔法原子. A separate wave of new capture devices appeared on the floor from 鹿明机器人, 帕西尼, 度量科技, and 灵御智能. The mechanism is straightforward. The ceiling on what these models can now do is gated less by architecture than by the quality and quantity of physical-action training data. Owning the collection pipeline is the new moat.
The third trend is that dexterous hands are converging on human-hand specifications. Five fingers, roughly 1:1 human-hand scale, around 20 degrees of freedom, integrated tactile sensors (触觉传感器), and direct-drive (直驱) or hybrid actuation have become the de facto hardware template on the show floor. The driving logic is that motion captured from human demonstrations transfers more cheaply to a robot hand of the same scale, and a hand that looks like the human hand it learns from requires far less translation by the model.
None of these three patterns is exclusive to China. They reflect what happens when an AI-hardware sector is too young to have a mature division of labor. Specialist robotics suppliers in more mature markets exist because decades of iteration built a reliable supply chain to specialize against. In fields where that supply chain does not yet exist, full-stack ownership is less a national strategy than a survival response.
The Leiphone piece itself flags the same caveat: the industry is early and lacks a mature standardized division of labor. That note cuts against reading the trend as a national-leadership story. Some of the full-stack push is genuine strategic vision. Some is companies filling component gaps they cannot yet buy off the shelf. The two forces are easy to confuse on a trade-show floor, where polished booths make surviving bets look like planned roadmaps.
An adjacent signal from the same Leiphone reporting reinforces the timing. 宇树科技 (Unitree), one of China's most visible humanoid-robotics firms, reportedly cleared Shanghai's STAR Market IPO review in 73 days from acceptance, with a reported 4.2 billion yuan raise and founder Wang Xingxing retaining roughly 70 percent voting control. A first A-share listing for an embodied-AI company would give the full-stack thesis a public-market test, though the IPO numbers cited here are Leiphone's own reporting and have not been independently verified by this desk.
What to watch next: whether any of the Chinese full-stack players offload a layer back to a specialist supplier once the supply chain matures, and whether a non-Chinese full-stack entrant, likely American, emerges in the next 12 to 18 months. The pattern travels. The geography may not.