Deep Reinforcement Learning for Surgical Robotics with State and Image Information: A Survey

Abstract Surgical robotics has become a cornerstone of modern minimally invasive procedures, offering enhanced precision, dexterity, and ergonomics compared to conventional manual techniques. As the field progresses toward (semi-)autonomous operation, learning-based methods, particularly reinforcement learning (RL), have the potential to endow surgical robots with adaptable, data-driven decision-making capabilities. However, despite notable success of deep learning in tasks such as tool and anatomy segmentation, the application of RL to surgical robotics has remained largely confined to simplified tasks. As a result, the exploration of RL for realistic, high-fidelity surgical scenarios remains sparse, leaving substantial room for methodological and clinical advancement.This survey presents a comprehensive review of deep RL methods for surgical robotics across domains including laparoscopy, endoscopy, ophthalmology, and related specialties. We categorize existing policy-learning approaches into seven principal areas, analyzing their applications, effectiveness, and limitations. We additionally review the surgical environments and simulators that currently or may support RL research, highlighting the physical phenomena they model and the implications for downstream policy transfer. Public datasets relevant to RL agent training and surgical scene understanding are also summarized, emphasizing their role in enabling reproducible research and data-driven skill acquisition.We conclude by discussing open challenges and emerging research directions critical to advancing RL-driven autonomy in surgical robotics. Our goal is to provide a structured map of the field and a clear perspective on future opportunities toward safe, robust, and clinically meaningful autonomous surgical systems.