부산대학교 도서관

In this paper, we applied reinforcement learning algorithm to automated counter traction control in surgery. Counter traction is a surgical technique where the forceps are manipulated so that tissue in surgical field is visible and positioned suitable for operation. Since this surgical procedure is performed repeatedly during surgery, automation is expected to reduce the burden on the surgeon. We incorporated position-based dynamics of point clouds simulating two-dimensional membrane tissue and reward function expressing optimal shape of tissue in terms of visibility and flatness with a reinforcement learning model that learns the optimal traction direction. The model was trained in a simulation environment to generate movement of forceps tips to realize counter traction. The results suggest that the desired manipulation could be acquired under several fixed conditions.