Toward Autonomous Soft-Tissue Minimal Invasive Surgery: Integrating Learning, Reasoning, Perceiving for Adaptive Surgical Robot Autonomy

This thesis investigates the foundations required to achieve adaptive autonomy in soft-tissue surgical robotics. Soft-tissue procedures are defined by continuous, unpredictable anatomical deformation, making autonomous execution dependent on the integration of learned surgical skills, real-time perception, adaptive reasoning, and precise robotic action. To address these challenges, this work develops a unified framework based on four interconnected pillars—Learning, Reasoning, Perceiving, and Acting (LRPA)—and demonstrates its feasibility on clinically inspired phantom scenarios. The first pillar, Learning, focuses on extracting expert surgical competence directly from demonstrations. A domain-adaptive imitation learning framework is developed to recover fine tool trajectories, orientations, and task states for tumor cutting and tissue manipulation. Policies are trained first in high-fidelity simulation, then transferred to real silicone phantoms, achieving an average positional error of 0.2 cm and generalizing to novel anatomical variations. The second pillar, Reasoning, establishes the geometric and anatomical logic needed for autonomous decision-making. This thesis introduces a complete reasoning pipeline for selective vascular clamping in partial nephrectomy, including the extraction of clamping candidates along the arterial tree and the formalization of surgical target points that must remain valid despite intraoperative anatomical motion. The third pillar, Perceiving, provides the system with real-time awareness of soft-tissue behavior. A fast, RGB-D–based deformation tracking method is developed to estimate dense non-rigid motion on organ surfaces. The resulting deformation field enables continuous updates to surgical targets, trajectories, and anatomical landmarks.​ Validation on CT-derived phantoms demonstrates robust performance under external perturbations and visually challenging conditions. Finally, Acting is enabled through the design of a custom actuated surgical tool capable of controlled cutting, grasping, and clamping. Integrated with the perception and reasoning modules, the system performs deformation-aware tumor cutting and dynamically updated path following, establishing the first demonstration of adaptive execution on deformable phantom organs. Together, these contributions create a cohesive framework for perceptually informed, adaptively controlled surgical robotics. The results outline a clear path toward safe, intelligent autonomy in soft-tissue surgery and provide essential building blocks for future deployment in real clinical environments.