China's physical-AI development is gaining ground as embodied-intelligence firms explore cross-embodiment scheduling - different-shaped robots sharing one underlying model. At one Beijing embodied-AI company, robots for convenience-store service, industrial handling, tennis and street dance all plug into the same embodied model, with agents letting training shift from high-precision data collection to learning directly from human-work videos.
At another Beijing AI firm, a dual-arm robot accesses a domestic open-source model as a skill pool, picking the right plan to grasp items even when blocked by a box, with inference several to dozens of times faster than the open-source baseline. A humanoid at a Guangzhou expo used a fast-brain plus slow-brain architecture to multi-task - fetching ice cream while waiting for coffee. Firms report 2026 embodied-intelligence compute demand several times that of 2025, as data and network infrastructure for physical AI are being built out.
