Over the past year, embodied AI news has come thick and fast. According to incomplete statistics from QbitAI, as of June 12, 2026, financing in China’s embodied AI sector reached about RMB 43.8 billion. Data from the State Taxation Administration shows that from January to May 2026, total purchases of embodied AI robots by industrial enterprises nationwide rose 2.3 times year on year. Corporate capital is shifting from “buying demos” to “deploying on production lines.”
Industrial scenarios are actually easier to commercialize before consumer scenarios: pick-and-place, transport, insertion and removal, and similar actions are highly repetitive, with relatively fixed task boundaries. Unlike home environments, they are not constantly disrupted by variables such as pets or visitors. But traditional industrial robots face two walls—fixed hardware structures mean that changing workstations requires redesign, while software relies on preprogramming and cannot understand scenes or instructions.
To break down these walls, a robot must first understand the scene, then execute actions accurately, and run stably for long enough without human supervision. Anu Intelligence chose six models to cover data generation, cognitive understanding, training evolution, action execution, and end-to-end deployment: the long-video generation model Helios achieves 19.5 FPS inference on a single H100 AI chip; SimLab