SurgLAT Enhances Robotic Laparoscope Control with Latent Attention

Autonomous control of laparoscopic cameras in robotic surgery is a complex challenge, primarily because the surgeon's focus, or "operative intent," is a dynamic and evolving latent state rather than a fixed physical object. This research introduces SurgLAT (Surgical Latent Attention Tracking), an innovative causal online framework designed to model this latent surgical attention and enable autonomous laparoscopic view control. SurgLAT employs a frozen DINOv3 encoder and a state-conditioned spatial token mixer to extract relevant operative evidence, guided by a memory-based spatial prior. A key component is its selective causal latent memory module, which effectively models both short-term motion continuity and long-term surgical intent by dynamically retrieving current, recent, and historical latent states. This learned latent surgical attention is then translated into a probabilistic heatmap and operative region for guiding the endoscope. Beyond perception, the framework includes a robotic deployment system that incorporates explicit Remote Center of Motion (RCM) constrained control, based on a virtual-axis formulation. This, combined with redundancy-aware null-space initialization, ensures stable and smooth manipulator motion. The full SurgLAT system has been validated using real laparoscopic surgical videos and a physical robotic laparoscope platform, demonstrating robust online tracking and stable autonomous adjustments even under challenging conditions like occlusion, rapid motion, and target transitions.