On October 26th, a research report by CITIC Securities highlighted that the application scope of general large-scale models is broadening, transitioning from high-level task planning for robots to encompassing action generation and execution control. Currently, robotic intelligence predominantly hinges on specialized models like VLA (Vision-Language-Action) and WAM (World Action Models), which are trained on robot action datasets. These models excel at translating visual and linguistic instructions into precise actions.

Recently, pioneering research institutions, including NVIDIA and Stanford University, have embarked on exploring the integration of GPT-6 Astra into robotic control systems. Their objective is to harness the formidable capabilities of general models in visual comprehension, spatial reasoning, and task planning, thereby diminishing the dependency on specialized policy training. Notably, Astra has showcased remarkable generalization abilities across embodied task datasets.

CITIC Securities forecasts that by 2027, the infusion of general models into the realm of embodied tasks will emerge as a novel and promising investment avenue within the AI industry. Moreover, in the near term, it is more plausible that general models and specialized models will coexist in a collaborative framework, each contributing to their respective strengths.