MS Thesis Defense: Cameron Wolfe

Optional ZOOM LINK

"Learned Modeling and Control of Continuum Robots for Minimally Invasive Surgeries"

Abstract

To reduce morbidity and mortality during surgery, surgeons have increasingly turned to minimally invasive surgery (MIS), which involves passing instruments through small incisions or natural orifices to minimize patient trauma. Although MIS has significantly improved patient outcomes, it hinders a surgeon's dexterity by impairing visual and tactile feedback. These deficiencies have prompted the adoption of robot-assisted surgery (RAS), in which surgeons control robots instead of using handheld instruments. While RAS has improved patient outcomes, robots struggle to navigate constricted spaces due to their rigidity, spurring the development of continuum robots (CRs). These flexible infinite degree-of-freedom robots move by bending, allowing them to snake through passageways without damaging surrounding tissue. However, this flexibility complicates modeling and control due to the nonlinear effects of material properties, external loads, and material degradation on movement.

I have continued the work of Carolina Lago, Pena Maia, and Brook Leigh in developing a low-cost CR for surgical applications. My work has involved upgrading the robot by creating a new, more resilient, flexible spine, reducing the robot's frame volume from 5.86 to 0.66 liters, developing hardware to improve tip pose accuracy measurement, and writing robust control and communications software. With this enhanced robot, I used an artificial neural network (ANN) to learn an accurate, lightweight forward kinematics model of the robot based on measured input/ output data. The conventional modeling approach, known as constant curvature modeling, assumes the robots bend in a series of arcs, which makes modeling tractable but significantly degrades accuracy. My learned model outperforms existing constant curvature models, with an RMS position error of 2.2 mm vs 9.5 mm on our robot. I then integrated this ANN-based model into optimal open-loop and Jacobian-based closed-loop control policies. The robot uses these control policies to follow a representative trajectory through the robot's workspace, with the open and closed-loop policies achieving RMS position errors of 2.415 mm and 1.619 mm, respectively. My work demonstrates the potential of learned kinematic models for developing accurate and reliable CRs, which have the potential to lower costs and improve patient safety in the operating room.

Thesis Committee

• Prof. Ryan Halter (chair)
• Prof. Mike Kokko
• Prof. Minh Phan

Contact

For more information, contact Thayer Registrar at thayer.registrar@dartmouth.edu.