SRT-H: A hierarchical framework for autonomous surgery via language-conditioned imitation learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Surgical Technology

Background:

• Current autonomous surgery research is limited to simple tasks in controlled settings.
• Real-world surgery requires dexterity, long-term operation, and adaptability to tissue variability, which existing methods struggle with.

Purpose of the Study:

• To develop a hierarchical framework for dexterous, long-horizon autonomous surgical procedures.
• To enable robots to generalize to the variability of human tissue and recover from errors.

Main Methods:

• A hierarchical framework with a high-level task planning policy and a low-level trajectory generation policy.
• The high-level planner uses language space for task instructions and error correction.
• Validation through ex vivo cholecystectomy experiments and ablation studies.

Main Results:

• Achieved a 100% success rate in autonomous ex vivo cholecystectomy on eight gallbladders.
• Demonstrated improved error recovery from suboptimal states in dynamic surgical environments.
• The hierarchical approach enhanced policy robustness for realistic surgical applications.

Conclusions:

• The proposed framework enables step-level autonomy in surgical procedures.
• This research marks a significant advancement toward the clinical deployment of autonomous surgical systems.
• The language-guided hierarchical approach is effective for complex, long-horizon surgical tasks.