AI-enhanced intraoperative decision-making in robotic surgery: from augmented reality navigation to foundation models

Robotic surgical platforms have improved instrument dexterity, tremor filtration, and ergonomic control, yet intraoperative decisions still depend on a surgeon interpreting a moving, partially observed field under time pressure. Artificial intelligence (AI) is increasingly used to close parts of this loop by recognizing anatomy, tracking instruments, anticipating events, and-experimentally-planning or executing selected actions. This review maps the computational technologies that support intraoperative decision-making in robotic surgery and appraises the clinical evidence for them.Because the direct robotic evidence base is still uneven, we explicitly distinguish three kinds of material: (i) studies conducted on robotic systems during surgery; (ii) non-robotic surgical evidence from laparoscopic and thoracoscopic procedures that use the same vision and workflow methods; and (iii) preclinical, preoperative, and perioperative work that is conceptually relevant but not intraoperative. Robotic and non-robotic settings are labeled throughout the tables and text. We further separate non-AI image guidance-indocyanine green (ICG) fluorescence and registered overlays-from AI-based inference, and we include image-guided endovascular robotics as a complementary setting in which fluoroscopic perception, vascular-topology reasoning, and closed-loop instrument control are particularly advanced.We organize the field into four computational layers: perception, anatomical and procedural reasoning, action planning, and control and execution. Augmented-reality and fluorescence navigation can improve orientation and selected intermediate outcomes, and several randomized trials now exist. Computer-vision systems can recognize anatomy, phases, instruments, and adverse events, and in endovascular work they can track guidewires and reason over vascular topology. Foundation models broaden language and multimodal capabilities but remain, in surgery, largely preclinical. Autonomous subtasks have shown promise in phantoms, animals, and a small number of carefully bounded tissue tasks; no system has demonstrated safe, accountable patient-level autonomous decision-making.Navigation and AI-assisted systems have improved orientation, reduced ischemia, and accelerated recognition of critical events in some studies. Whether these intermediate gains improve patient outcomes, generalize across institutions, or justify their cost remains unknown. The question is not whether AI can interpret surgical scenes, but which decisions can be delegated, with what evidence, uncertainty limits, oversight, and outcome measures.