Expecting a glimpse of the frontier, the author instead found the World Robot Conference dominated by three archetypes: conventional robotic arms wrapped in humanoid shells, silicone figures animated by a chatbot, and generic components rebranded as robot-grade. The skepticism extends well beyond one show: a strategy consulting chief told media that most humanoids are essentially remote-controlled arms in a human shape, while Morgan Stanley figures show roughly 65% of first-half global robot shipments went to demos, reception, education and data collection rather than real production work.
The article's core argument is that physical AI only works when a trainable model and a real-world data flywheel exist together. Driving is the sole application that qualifies: millions of production cars continuously return road footage and steering, throttle and brake actions, and VLA-type models learn general physical understanding rather than just driving. Household tasks, by contrast, need a third data tier with force and tactile information, which today must be gathered machine-by-machine through prohibitively expensive teleoperation. That explains the great pivot — Tesla has Optimus, and at least eleven Chinese automakers are building robots or have set up robotics companies.
XPeng is presented as the clearest example. It has repositioned itself as a global embodied-intelligence company, and its second-generation VLA update can retain thirty seconds of recent road context, forecast six seconds of surrounding traffic movement and parse natural-language commands — capabilities the company says can migrate from cars to cockpits and robots. Its robotics arm just raised $900 million at a post-money valuation above $6.3 billion, and He Xiaopeng has noted that more than 85% of the IRON humanoid's supply chain overlaps with the car business. On the supplier side, Horizon Robotics — which founder Yu Kai says only Tesla can match in combined chip and algorithm capability, earning it a reputation as a Tesla without the carmaking — has reached over 100 robot categories and 400-plus customers through affiliate Dizua Robotics, covering more than half of China's embodied-AI firms, with a next-generation foundation model fusing world models promised within six months.
The verdict: smart-driving companies are becoming the substrate of the robotics industry, with the author likening autonomous driving to a student near 95 out of 100 while embodied intelligence sits between 20 and 60. The missing piece is force data, which remains expensive and controlled by no one — until that changes, the arms and dolls on display will depend on whoever upgrades the chatbots inside them.