A Novel Model-Independent Approach for Autonomous Retraction of Soft Tissue

Surgical robots have played an important role in guiding surgery toward precise interventions and targeted therapy. Aim to provide standard surgical solutions, autonomous surgical robotic systems can potentially improve the reproducibility and consistency of robot-assisted surgery. To meet the demands of enhancing the autonomy of surgical robots, we propose a novel model-independent training methodology for retraction and manipulation of deformable tissue in robotic surgery. We use contrastive optimization to learn deformable tissue’s underlying visual latent representations. Then we apply an improved version of Deep Deterministic Policy Gradients (DDPG) with asymmetric inputs to train an agent in a simulation environment. To evaluate the effectiveness and validity of our method, we tested our approach based on three different criteria. Compared to the state-of-the-art method, our method could especially accomplish safe tissue retraction task in constrained situations without collecting expert demonstrations.