SurgVLA-Bench Evaluates Vision-Language-Action Models for Surgical Robotics

While Vision-Language-Action (VLA) models show promise for embodied intelligence, a standardized evaluation platform for surgical robotics has been lacking. This research addresses that gap by presenting SurgVLA-Bench, the first comprehensive benchmark specifically designed for laparoscopic surgical robotics. It utilizes the SurRoL simulation platform to create a hierarchical task taxonomy, ranging from basic atomic actions to complete surgical procedures. The benchmark employs a multi-dimensional evaluation framework that assesses both the accuracy of actions and the semantic consistency of the VLA models' understanding. The researchers systematically evaluated two main paradigms: autoregressive models (like OpenVLA) and flow matching models (like SmolVLA). Results indicate that autoregressive models generally excel in semantic understanding, while flow matching models often achieve higher task precision. However, even the best-performing models still face significant challenges. The inherent physical bottlenecks of laparoscopic surgery, such as a constrained endoscopic field of view, limited viewing angles, and frequent occlusions, remain fundamental hurdles that current VLA models struggle to overcome satisfactorily.