Surgical Robot Learning: From Demonstration and Simulation to World Models-A Review

摘要

Robot-assisted minimally invasive surgery is widely adopted, yet systems are predominantly teleoperated, and progress towards autonomy remains challenging. Expert demonstrations are limited and costly, while existing simulators struggle to capture contactrich tissue dynamics and intraoperative variability, contributing to sim-to-real gaps. This review surveys over 100 studies on surgical robot learning for manipulation tasks, organised by learning paradigms, data strategies, and model architectures. We find that, despite notable advances-including strong perception performance and demonstrations of complete ex vivo procedures-scaling remains limited by data scarcity and restricted environmental diversity. We discuss Surgical World Models (SWMs) as one possible direction: models that learn environment dynamics from endoscopic video and robot trajectories and can be used to synthesise large, diverse, and parameterizable training scenarios. Such approaches could complement physics-based simulation and targeted data collection; their practical value will depend on rollout horizon, fidelity under distribution shift, and verification and validation requirements for safety-critical use.