Five embodied AI foundation models were released within eight days, and Juxie Intelligent closed a round above RMB 100 million with Ecovacs and Zibianliang participating. The two facts sit at opposite ends of the same stack: model supply and hardware supply.
The release cadence indicates the route debate has narrowed toward multimodal foundation models. When five teams converge on comparable architectures inside one news cycle, differentiation shifts to data pipelines, evaluation methods and deployment cost rather than model category. Joint modules follow the same logic. The Juxie round places actuators and joints as the contested layer of humanoid hardware, with Ecovacs supplying manufacturing and supply chain depth.
Data requirements are being repriced downward. TwinDEX completed a 24-step chemistry experiment without real-robot teleoperation data in post-training, which weakens assumptions built into teleoperation-heavy collection programmes. In parallel, UBTECH validated factory layout, logistics and process flow in simulation before construction, evidence that digital twins now compress physical build cycles.
Commercial positioning is moving outward. Astribot presented the T1 at IROS 2026 with a starting price of US$18,000, giving overseas integrators a comparable price anchor for Chinese humanoid products. Patent attorney Dominic Davies argued in The Robot Report that intellectual property portfolios, not laboratory results, will decide the physical AI race, urging startups to secure core inventions before competition intensifies.
Two observations follow. First, model release density and hardware funding are converging on one bottleneck: verified task data and the modules that execute it. Second, a public US$18,000 price and an explicit patent argument appearing together suggest the criteria of judgement are shifting from demonstration capability to procurement feasibility and legal defensibility.