Learning based MPC for tissue manipulation

Robot-assisted surgery (RAS) has revolutionized surgical procedures, offering advantages over traditional methods. Currently, surgeons control robots for tasks like suturing and organ resection, leading to physical strain while the robot helps at the lower level by scalling down the movement and reducing hand tremor. These tasks are intricate and demand high dexterity and skill, leading to physical and cognitive strain for surgeons when performed manually. Automating these tasks presents an opportunity to improve surgical outcomes, mitigate human error, and reduce strain on surgeons.
Various control algorithms, including model-based (e.g., model predictive control) and model-free (e.g., reinforcement learning) approaches, have been developed. Model based control ensures sfety and stability however, their relience on accurate model of the system poses challanges in dynamic surgical settings. Model-free, specifically RL based control offeres adaptability by learning form the environemnt, however they requiers extensive interaction with the environemnt posing problems of safety and stability specifically during the initial phases.
To overcome these limitation we have porposed an (event-triggered) learning MPC for automating tissue manipulation. The proposed control uses accumulated error to trigger the Gaussian process and update the tissue model, thus adapting to changing dynamics online, while ensuring the safety and stability by imposing constraints.

Faculty Supervisor:

Mahdi Tavakoli

Student:

Partner:

University of Naples

Discipline:

Engineering

Sector:

Education

University:

University of Alberta

Program: