According to CoinMeta, a research report by CITIC Securities states that GPT-6 and Astra are entering the field of robot control, marking a significant turning point for physical AI. General-purpose large models are moving beyond high-level task planning for robots to also encompass action generation and execution control. The "brains" of robots mainly rely on specialized models such as VLA and WAM that have been trained with robot action data to convert visual and linguistic commands into actual actions. Recently, research institutions like NVIDIA and Stanford have begun exploring the application of GPT-6 and Astra in robot control, attempting to leverage the existing visual understanding, spatial reasoning, and task planning capabilities of these general-purpose models to reduce reliance on specialized strategy training. Astra has demonstrated good generalization on datasets for embodied tasks. CITIC Securities believes that looking ahead to 2027, the entry of general-purpose models into the field of embodied tasks will become the next major investment direction for the AI industry, and in the short term, there is a higher likelihood of a collaborative division of labor between general-purpose and specialized models.