Goldman Sachs' recent mid-year research report on Chinese humanoid robots indicates the domestic embodied AI industry is shifting from discussions of single VLA models towards multi-modal architectures that fuse VLA/VTLA with world models.

The report posits that world models are becoming a crucial functional layer alongside action models, enhancing a robot's next-state prediction, action verification, planning capabilities, and robustness in the real world.

The report mentions that companies clearly defining their next direction as combining VLA/VTLA with world models include ONEROBOTICS (06600), Galaxea, and Galbot.

ONEROBOTICS is the only Hong Kong-listed company among those mentioned, and its home-focused embodied AI strategy has become a significant point of observation in the report.

According to Goldman Sachs, ONEROBOTICS's target architecture is an end-to-end embodied AI system, transitioning from a modular system towards a fusion of VLA and world models.

Its OneModel 1.7 employs a latent world action model, integrating the generalization power of world models with the execution capabilities of VLA, and adds motion-centric control and success memory layers tailored for household tasks.

Notably, compared to relatively structured environments like industrial and logistics settings, home environments have long been considered a more complex and scarcer source of data for embodied AI.

The Goldman Sachs report notes that ONEROBOTICS management emphasizes the structural shortage of home environment data, positioning its Qianhai data factory as a dedicated home-scenario data center covering real-world multi-modal tasks such as laundry, cooking, and dishwashing.

Regarding commercialization, the report indicates ONEROBOTICS is currently focusing on semi-structured service scenarios like retail and elderly care, followed by data collection and validation through government deployments, aiming for consumer-grade product rollout on an approximately three-year cycle.

Its latest general-purpose humanoid robot, Onero, targets service scenarios that are easier to deploy, and the company views a Bill of Materials cost of around $3,000 as a key threshold for household adoption.

Industry observers believe that as embodied AI moves from technical demonstrations to engineering validation, the competitive focus for robotics companies will shift beyond just hardware platforms to who can first establish a closed-loop of "real-scenario data — model training — product deployment feedback."

For ONEROBOTICS, the iteration of OneModel, accumulation of home-scenario data, progress in the scaled deployment of Onero, and its cost reduction trajectory are likely to become key variables for subsequent capital market observation.