Photo credit: www.sciencedaily.com
Advancements in Robotic Surgery Through Imitation Learning
Groundbreaking research has shown that a robot, trained for the first time using videos of experienced surgeons, can perform surgical procedures with a proficiency comparable to that of human doctors.
This innovative approach, which utilizes imitation learning for training surgical robots, suggests a significant shift in how robotic surgery could be conducted. By removing the need to individually program every movement required for a surgical procedure, this method paves the way for robotic systems that can operate with greater autonomy, potentially allowing them to tackle complex surgeries independently.
“It’s truly remarkable to see this model excel when we provide it with input from cameras; it can accurately predict the necessary robotic maneuvers for surgery,” stated Axel Krieger, the senior author of the study. “We believe this represents a major leap forward in the realm of medical robotics.”
The recent findings, led by researchers from Johns Hopkins University, are set to be featured at the Conference on Robot Learning in Munich, which is renowned for its focus on advancements in robotics and machine learning.
In collaboration with researchers from Stanford University, the team trained the da Vinci Surgical System robot to perform essential surgical tasks such as needle manipulation, tissue lifting, and suturing. The model uniquely merges imitation learning with a machine learning structure similar to that used in ChatGPT. However, rather than processing text, this model interprets “robot” movements through kinematic calculations that quantify robotic motion into mathematical terms.
The researchers utilized a vast database of recordings from wrist cameras attached to da Vinci robots, capturing footage during actual surgical procedures conducted by global surgeons. This extensive repository of nearly 7,000 da Vinci robots provides a rich source of data, allowing the robots to learn through imitation. More than 50,000 surgeons have undergone training on this system, further contributing to the archive.
While the da Vinci system is prevalent in surgical settings, its precision has often been criticized. However, the research team found an effective method to address these inaccuracies by instructing the model to focus on relative movements rather than absolute actions, which can be prone to error.
“With just image input, this AI system can determine the appropriate action needed,” explained Ji Woong “Brian” Kim, the lead author. “We discovered that even with only a few hundred demonstrations, the model successfully learns the procedures and adapts to new environments it hasn’t previously encountered.”
During their training regimen, the robot adeptly learned to execute the three core tasks, performing each with a level of skill akin to that of human surgeons. Krieger highlighted the model’s impressive capability, noting, “What’s fascinating is that it explores actions we haven’t explicitly taught it. For instance, if it drops the needle, it intuitively knows to pick it up and proceed with the task, which is not something I programmed into it.”
The potential applications of this model extend beyond basic tasks. Researchers anticipate that the model could facilitate the rapid training of robots for a wide variety of surgical operations. Currently, the team is exploring the use of imitation learning for training robots to conduct full surgical procedures rather than just isolated tasks.
Historically, programming robots for surgical functions involved extensive hand-coding, often taking years to refine processes such as suturing, and that was just for a single type of surgery, according to Krieger. He emphasized, “This represents a significant change; now, we only need to gather imitation data from a variety of procedures, allowing us to train robots within a few days. This advancement accelerates our journey toward achieving autonomy in robotic surgery, minimizing medical errors, and enhancing surgical precision.”
Contributors from Johns Hopkins University include PhD student Samuel Schmidgall, Associate Research Engineer Anton Deguet, and Associate Professor of Mechanical Engineering Marin Kobilarov. From Stanford University, PhD student Tony Z. Zhao also contributed to this innovative research.
Source
www.sciencedaily.com