GF SEC: After Large Language Models, When Will Embodied Models Have Their GPT Moment?

GF SEC released a research report stating that after crossing the mass production threshold, the focus of industry development has shifted to embodied models. The industry has now passed the mass production node, hardware solutions are basically finalized, and the development focus has gradually migrated from hardware-layer components to software-side embodied modelswhich affect the level of intelligence and motion execution performance of embodied intelligence. Their model capabilities also determine the application scenario boundaries and terminal demand of embodied intelligence, making them the key variable determining the pace and stage of subsequent industry development. GF SEC's main views are as follows: Embodied models: In the pre-GPT era, multiple routes are developing in parallel, advancing through the fog, with a large gap remaining from the ideal goal Just like the large model industry after the Transformer was proposed in 2017 and before GPT converged: multiple routes in parallel, winners undecided, advancing through the fog. Current embodied models are attempting three routes to achieve Siasun Robot&Automation's ideal intelligent capabilities of "stability, accuracy, strong generalization, and cross-scenario" application: (1) Hierarchical brain-cerebellum: the brain large model is responsible for high-level planning + the cerebellum skill library executes specific actions; this is a solution with high maturity and relatively fast implementation for specific tasks and specific scenarios, but with a low capability ceiling and weak generalization ability; (2) VLA: reconstructs action representation and action generation, and is the current mainstream layout route, but long-duration tasks and out-of-distribution scenarios remain shortcomings; (3) World models: imagine first, then decide, then execute, the long-term form closest to "understanding the physical world," but still in its early stages. The convergence of embodied intelligence model routes requires the return of two major cornerstones Referring to the convergence history of large models, convergence requires two cornerstones. (1) The "three major elements": the data wallmainly referring to model training data; currently, embodied models lack massive standardized data, and training costs are high; the representation walllacking a unified task objective similar to "predicting the next word," model capabilities cannot accumulate across tasks, scenarios, and embodiments; the closed-loop wallphysical trial and error is costly and feedback is slow, the industry lacks a unified evaluation standard, and iteration cannot be self-driven. (2) The "signpost" of the convergence direction: large models began to converge only after the Scaling Law emerged under the premise of satisfying the three walls. At present, a convergence law similar to the Scaling Law for embodied models has not yet appeared, the correctness of routes cannot be verified, and the industry can only conduct dispersed trial and error across multiple lines. Domestic and international manufacturers are pursuing multi-route layouts to seek the ideal answer for embodied models Domestic manufacturers: (1) UBTECH ROBOTICS: simultaneous layout of multiple routes + industry vertical models, iterating from the first-generation Thinker foundation and Thinker-VLA to the open-source Thinker, Thinker-WM world model, and Walker embodied intelligence model. UBTECH ROBOTICS' embodied models balance technological generalization and full-machine engineering implementation; (2) Unitree Robotics: dual-track parallel development of the WMA world model and VLA. The world model UnifoLM-OminiA-0.3 launched in July 2026 mainly adopts the video prediction route and has verified its application potential for consumer-end scenarios; (3) Galbot: simultaneous advancement of three lines: hierarchical models, VLA integration, and world models. International manufacturers (1) Physical Intelligence: adheres to the VLA technology route, and 0.7 has become one of the current frontier VLA models; (2) Figure AI: focuses on the VLA hierarchical technology route and has achieved deployment in industrial and consumer scenarios; (3) NVIDIA: centered on foundation model R&D, launching the GR00T series and deeply binding the Cosmos world model with the full-machine ecosystem. Risk warnings Technology iteration falling short of expectations, impact of safety and privacy regulations, and ethical factors.