Shanghai's intelligent manufacturing action plan sets a target of 8,000 humanoid robots deployed in factories by 2028, against a base of 30 national-level smart manufacturing facilities already in the city. In the same news cycle, Unitree's Wang Xingxing identified under-trusting AI as his biggest past error and placed the embodied AI tipping point within one to two years.
Two platform launches show where the supply side is consolidating. JUNPING Intelligent released Primus Stack, bundling data collection, model training, skill development and production-line delivery; D-Robotics showed Sunrise chips, RDK developer kits and RoboGo software at CIIF, covering vision, data collection and embodied control. Product boundaries are shifting from single models toward packaged data pipelines and deployment tooling.
Capital moved into the weaker links of that chain. Matrix Innovation secured tens of millions of RMB from X Square Robot and XBOT PARK Fund for tactile sensors, a component that gates dexterous hand performance; Xianyin Future closed three rounds in four months, raising hundreds of millions of RMB for space robotics. Allocation is spreading beyond the humanoid body itself toward sensing and non-factory scenarios.
Field evidence is setting a stricter bar. Boston Dynamics reports Spot handling SWAT, EOD and Hazmat tasks, and one assessment frames laundry as the actual readiness test. Sprinting and boxing demos read as capability claims; manipulation reliability under unstructured conditions remains the binding constraint.
Two observation points follow. Tactile sensor cost and durability determine whether dexterous manipulation scales beyond pilots, and city-level deployment targets need to convert into procurement orders before 2028 to count as demand; otherwise they remain planning documents. Until then, full-stack delivery capacity explains more than model benchmarks.