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

Diego Dall’Alba, Michał Nasket, Sabina Kaminska, Przemysław Korzeniowski

Robotic-assisted surgery is evolving at a fast pace and holds significant potential for improvements through automation. Yet, methods like Reinforcement Learning (RL), which require extensive task repetition, are challenging to apply directly in real surgical scenarios due to safety and feasibility concerns. This highlights the importance of using simulated environments that combine realism with computational efficiency and scalability.
In response to this need, we present FF-SRL (Fast and Flexible Surgical Reinforcement Learning) — a high-speed, GPU-based simulation platform tailored for robotic surgery. Unlike traditional setups, FF-SRL runs both the physics-based simulation and the RL training process entirely on a single GPU. This design eliminates common performance limitations caused by data exchange between the CPU and GPU, significantly boosting learning speed.
Experimental results demonstrate that FF-SRL can decrease the training duration for intricate tasks like tissue manipulation by approximately tenfold — achieving performance in just a few minutes compared to conventional hybrid simulators. This level of efficiency opens new possibilities for testing and refining RL algorithms in surgical contexts. To support further research and collaboration, we have made the FF-SRL framework freely accessible to the research community.

Authors: Diego Dall’Alba, Michał Naskret, Sabina Kaminska, Przemysław Korzeniowski

DOI: 10.1109/IROS58592.2024.10801658

Keywords: Robotic-assisted surgery, Reinforcement Learning (RL), Surgical simulation, GPU-based simulation, Fast and Flexible Surgical Reinforcement Learning (FF-SRL), Real-time physics simulation, Computational efficiency, Simulation platform