基于强化学习的机器人眼科手术适应性虚拟夹具设计

Design of adaptive virtual fixture for robot-assisted ophthalmic surgery based on reinforcement learning

虚拟夹具在提高人机协作手术的操作性能和安全性方面发挥着重要作用。然而,虚拟夹具一般是预先人工设计的,无法适应操作者指定的技能。基于对手术数据的学习,本文提出了一种基于强化学习策略训练方法的机器人辅助眼科手术虚拟夹具控制的通用框架,来辅助医生完成手术任务。虚拟夹具的执行模块经过了重新设计,用来增强顺应性阻抗而不是固定阻抗,并且在从端生成了动态笛卡尔力反馈。技能学习模块基于强化学习方法进行虚拟夹具的调整以及配置优化。本文提出的强化学习策略训练方法的有效性通过模拟白内障撕囊实验进行评估。在所提出方法的帮助下,经验丰富医生的操作精度提高了50.3%,新手医生的操作精度提高了45.5%。在不花费更多执行时间的情况下提高了性能,这表明了所提出的方法使虚拟夹具的冗余减少,并且提高了虚拟夹具对于经验丰富医生和新手医生的适应性。

Virtual fixture plays a significant role in enhancing manipulation performance and safety for man-machine collaborate surgery.However,the virtual fixture is pre-fixed and hand-designed,unable to adapt to the operator-specified skills.Based on the learning of surgical data,a general framework for virtual fixture control of robot-assisted ophthalmic surgery based on reinforcement learning policy training method is proposed to assist doctors to complete surgical tasks.The virtual fixture enforcement module is redesigned to reinforces compliant impedance rather than a fixed one and generates dynamic Cartesian force feedback on the client side.The skill learning module is carried out the virtual fixture adjusting and configuration optimization based on reinforcement learning method.The proposed method is evaluated by simulated cataract capsulorexis experiments.With the assistance of the proposed method,the performance of experienced doctors is enhanced in precision by 50.3%,to fresh doctors the precision is increased by 45.5%.The performance is enhanced without spending more execution time,which shows the proposed method enables the cutback of virtual fixture redundancy and the improvement of adaptability of virtual fixture to both experienced and fresh doctors.