CITIC Securities report: GPT-6 Astra moves into robot control

On October 26, 2026, at 08:59, 36Kr reported on a research note from CITIC Securities claiming that general-purpose large models are extending from high-level robotic task planning into motion generation and execution control. According to the note, robot brains today rely mainly on specialized models such as VLA and WAM, which are trained on robotic motion data and convert vision and language instructions into actions. CITIC Securities added that research organizations including NVIDIA and Stanford have recently begun exploring the use of GPT-6 Astra for robotic control, drawing on the general model's existing capabilities in visual understanding, spatial reasoning and task planning in order to reduce dependence on specialized policy training. The note further stated that Astra has shown relatively strong generalization on embodied-task datasets, and projected that by 2027 the entry of general models into embodied tasks will become the AI industry's next major investment direction, while in the near term general models and specialized models such as VLA and WAM are more likely to form a division of labor and work in coordination. This is the only publicly available information about the note at present.