Chinese Robots Are Dancing, American Robots Are Publishing Papers

Robots on stage, robots in journals

Chinese humanoid robots have become a pop-culture phenomenon — from Spring Festival Gala lineups to factory-floor demonstrations — while U.S. efforts are winning headlines in research journals and preprint servers. The contrast is stark. China’s manufacturers show off dexterity and scale. The U.S. showcases algorithms and world models. Which matters more: flashy demos or provable general intelligence?

Hardware versus software — a deliberate split

China’s strength is manufacturing. Companies and integrators from Xpeng (小鹏) to Li Auto (理想) and component suppliers such as Sanhua Intelligent Controls (三花智控) and Joyson Electronics (均胜电子) are retooling automotive supply chains for humanoid joints, batteries and thermal systems. It has been reported that some of these suppliers are in talks to supply parts for Tesla’s Optimus program; other supplier links remain rumors. By contrast, U.S. efforts — from DeepMind and OpenAI alumni to startups like World Labs (founded by Li Fei‑fei, 李飞飞) and Figure — focus on software: large “world” models, spatial intelligence and end‑to‑end learning. Figure’s VLA model and Tesla’s public “end‑to‑end” video learning push showrooms and white papers moving in tandem; reportedly these lab outputs have materially affected valuations and investor attention.

Supply chains, geopolitics and what comes next

The resemblance to the electric‑vehicle era is deliberate: both EVs and humanoid robots are “AI brain + actuator” systems that reuse batteries, motors and control electronics. But geopolitics matters. U.S. chip and export controls complicate access to the most advanced silicon, reinforcing Nvidia’s (英伟达) dominance with robot‑focused platforms like Isaac and Jetson Thor and accelerating domestic moves in China to secure supply chains. Does this produce complementary strengths — Chinese muscle, American mind — or a brittle bifurcation? Expect a long, competitive race where physical engineering and simulation‑grade software must meet.