Scholarlink

Purpose <p>Reliable automation of surgical suturing requires accurate and flexible simulation tools to effectively learn robot autonomous skills that match clinical practice. Yet, high-fidelity, ready-to-use learning frameworks remain scarce even for popular platforms like the da Vinci Research Kit (dVRK). We release Nail It!, a complete Unity–ROS–dVRK learning framework supporting both teleoperation data collection and Reinforcement Learning (RL) training of surgical primitives like needle grasping and placement.</p> Methods <p>Nail It! features (i) a physics-based Unity environment with full and accurate dVRK kinematic modelling, (ii) a real-time ROS communication with the dVRK master console for active surgeon control of Patient-Side Manipulators (PSMs), and (iii) an integrated Graphical User Interface (GUI) for environment tuning, reward design, learning algorithms development, and sim-to-real validation. We also provide full support for RL policies training, e.g. through Proximal Policy Optimization (PPO) with curriculum learning and domain randomization for robust policy training of suturing steps, and Multi Agent RL for multi-arm coordination.</p> Results <p>We experiment with Nail It! to learn autonomous suturing with our dVRK using PPO with curriculum learning to handle multi-step surgical procedures, and domain randomization techniques to mitigate the sim-to-real gap and enable policy deployment. Nail It! delivers policies that can solve the suturing task in simulation with 95% accuracy, with proven robustness over visual and positional variations up to 15&#xa0;mm (target).</p> Conclusions <p>Nail It! provides a modular, high-fidelity platform for developing and benchmarking autonomous surgical skills. By integrating accurate kinematics, reinforcement learning, and teleoperation, it supports both autonomous training and human-in-the-loop control in realistic surgical scenarios. Direct ROS connectivity and an integrated GUI further facilitate rapid development and sim-to-real experimentation. Supplementary materials associated with this work are publicly available at <a href="https://luigimuratore.github.io/Nail-it/">https://luigimuratore.github.io/Nail-it/</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Nail It! A learning framework for autonomous surgical suturing and teleoperation on the dVRK

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月2日

查看全文
  • Luigi Muratore,
  • Matteo Pescio,
  • Federica Barontini,
  • Francesco Marzola,
  • Pietro Leoncini,
  • Carlo Alberto Ammirati,
  • Alberto Arezzo,
  • Giulio Dagnino,
  • Giuseppe Averta

摘要

Purpose

Reliable automation of surgical suturing requires accurate and flexible simulation tools to effectively learn robot autonomous skills that match clinical practice. Yet, high-fidelity, ready-to-use learning frameworks remain scarce even for popular platforms like the da Vinci Research Kit (dVRK). We release Nail It!, a complete Unity–ROS–dVRK learning framework supporting both teleoperation data collection and Reinforcement Learning (RL) training of surgical primitives like needle grasping and placement.

Methods

Nail It! features (i) a physics-based Unity environment with full and accurate dVRK kinematic modelling, (ii) a real-time ROS communication with the dVRK master console for active surgeon control of Patient-Side Manipulators (PSMs), and (iii) an integrated Graphical User Interface (GUI) for environment tuning, reward design, learning algorithms development, and sim-to-real validation. We also provide full support for RL policies training, e.g. through Proximal Policy Optimization (PPO) with curriculum learning and domain randomization for robust policy training of suturing steps, and Multi Agent RL for multi-arm coordination.

Results

We experiment with Nail It! to learn autonomous suturing with our dVRK using PPO with curriculum learning to handle multi-step surgical procedures, and domain randomization techniques to mitigate the sim-to-real gap and enable policy deployment. Nail It! delivers policies that can solve the suturing task in simulation with 95% accuracy, with proven robustness over visual and positional variations up to 15 mm (target).

Conclusions

Nail It! provides a modular, high-fidelity platform for developing and benchmarking autonomous surgical skills. By integrating accurate kinematics, reinforcement learning, and teleoperation, it supports both autonomous training and human-in-the-loop control in realistic surgical scenarios. Direct ROS connectivity and an integrated GUI further facilitate rapid development and sim-to-real experimentation. Supplementary materials associated with this work are publicly available at https://luigimuratore.github.io/Nail-it/.