<p>The rapid development of generative artificial intelligence and large models, including large vision models (LVMs), has accelerated their wide applications in medicine. Robot-assisted surgery (RAS) or surgical robotics, in which vision has a vital role, typically combines medical images for diagnostic or navigation abilities with robots with precise operative capabilities. In this context, LVMs could serve as a revolutionary paradigm towards surgical autonomy, accomplishing surgical representations with high fidelity and physical intelligence and enabling high-quality data use and long-term learning. In this Perspective, vision-related tasks in RAS are divided into fundamental upstream tasks and advanced downstream counterparts, elucidating their shared technical foundations with state-of-the-art research that could catalyse a paradigm shift in surgical robotics research for the next decade. LVMs have already been extensively explored to tackle upstream tasks in RAS, exhibiting promising performances. Developing vision foundation models for downstream RAS tasks, which is based on upstream counterparts but necessitates further investigations, will directly enhance surgical autonomy. Here, we outline research trends that could accelerate this paradigm shift and highlight major challenges that could impede progress in the way to the ultimate transformation from ‘surgical robots’ to ‘robotic surgeons’.</p>