FF-SRL: High Performance GPU-Based Simulation For Surgical Robot Learning

Robot-Assisted Surgical Systems (RASS) are increas- ingly widespread and researchers aim to introduce au- tomation during operations, supporting a progressive increase in the autonomy of RASS [1]. Given the sig- nificant successes obtained in other robotic applications [2], many works adopt Reinforcement Learning (RL) to support the RASS execution of different tasks, such as the manipulation of deformable tissues [3], [4], and the removal of debris or liquids during operations [5]. Generally, RL training requires numerous task repeti- tions, which are unsafe and impractical in real RASS. Fast and Flexible Surgical Reinforcement Learning (FF- SRL) addresses this by creating a realistic and com- putationally efficient simulator that runs entirely on a single GPU, eliminating bottlenecks associated with data transfer. The goal is to accelerate learning rates, mak- ing experimentation with RL in robotic surgery more accessible. The results indicate a significant reduction in training time for complex tasks, down to a couple of minutes compared to conventional CPU/GPU simula- tors. Such advancements are important in facilitating the development of autonomous and efficient RASS through RL techniques.