Pretraining Using Comparable Human Activities of Daily Living Dataset in Robotic Surgical Task Learning

Abstract

Training robots to acquire surgical skills poses significant challenges, primarily due to the limited availability of comprehensive datasets and safety constraints that restrict real-time trial-and-error learning. Although human Activities of Daily Living (ADL) tasks differ substantially from surgical tasks, they encompass fundamental motor skills that can serve as a foundation for robot learning. Notably, skilled surgeons often develop their advanced surgical abilities by building upon these basic motor skills acquired through daily activities. Inspired by this progressive learning trajectory, we propose a novel surgical skill training framework that enables robots to learn basic motor skills from the ADL dataset and quickly adapt to advanced surgical skills. Specifically, we propose a unified predictive representation space, constructed using probabilistic successor features, which capture the dynamic patterns of motion primitives common to both ADL and surgical tasks. To investigate the transferability of skills from human ADL tasks to robotic surgical tasks, we conducted a mathematical analysis to evaluate transferable policies and performed simulation experiments to assess transfer performance. Furthermore, we validated the practicality and effectiveness of our method through real-world experiments. Results show that our method significantly reduces the need for extensive surgical datasets, and enables efficient learning in robotic surgical tasks.