Robot model learning and optimal control towards safer robotic surgery

Abstract

Surgery has experienced a large evolution over the past few decades, with the goal of reducing patients’ traumas as much as possible. Minimally Invasive Surgery (MIS) has been replacing open surgery thanks to the use of small incision on the patient’s body, thus limiting scar size, blood loss, and speeding up recovery.

The requirement of reducing invasiveness comes, however, at the cost of needing more complex and articulated devices, which are even harder to operate for surgeons. Robots can help overcome these limitations thanks to their control systems, allowing for a reduction of mental and physical burden for the surgeon. Additionally, the possibility to use robotic devices for surgical operations has sparked interest towards automating surgical procedures. Autonomous robots could perform a surgical task more precisely, with more dexterity than humans, and without being influenced by fatigue or tiredness.

Numerous are the challenges to solve before ensuring surgical procedures can be automated and properly performed by robots, such as sensing capabilities, visual perception and scene understanding, modelling and control. Surgical robots are becoming more and more complex due to the miniaturization requirements and the quest for high dexterity and flexibility, leading to highly articulated systems such as snake-like, soft, and continuum robots. Additionally, these requirements affect the choice of the actuation system, which is typically tendon-driven, and this, in turn, creates even more complications from the modelling and control perspective.

Whatever the level of automation, an important aspect is proper modelling and control of the surgical robots. In an environment where few (or no) external sensors can be used, controllers need to highly rely on the accuracy of the robot model and proper control strategies need to be developed. The research work presented in this thesis aims at investigating and proposing novel modelling and control techniques for surgical robots, in order to improve precision and safety in performing surgical tasks. Because of the complexities of the systems, analytical models are hard to compute and do not properly manage to properly describe such robots. The rising capabilities of Machine Learning (ML) approaches in building complicated models can be beneficial for modelling surgical robots. This thesis thus proposes novel methods based on Artificial Neural Network (ANN) for robot modelling and control.

Deterministic approaches are firstly investigated. A novel robust nonlinear regression approach based on ANN is proposed, capable of discarding outliers in the data. In addition, an ANN architecture leveraging differential relationships for model training is presented, showing how this approach can improve robot modelling and control. The proposed models are then used to develop control strategies capable of ensuring accurate path tracking under multiple task constraints.

One of the major complexities in surgical robots is caused by the nonlinearities in the tendon actuation. ML approaches should be able to inherently learn the compensation of tendon’s nonlinearities, but it is a challenging task. The thesis thus proposes two approaches for tackling backlash compensation. On one hand, the thesis proposes a method including a priori knowledge of the backlash model and compensation to help learn more accurate robot models; on the other hand, it proposes the use of advanced ML approaches such as Recurrent Neural Network (RNN) to assess how they can cope with those nonlinearities.

Despite their capabilities, ML approaches have the disadvantage of just relying on the training data, and thus being generally inefficient in generalizing to other tasks. The uncertainties in the data and in the model when reaching unexplored regions can lead to poor control performance and unstable behaviours. The thesis therefore introduces two control strategies based on probabilistic ML approaches capable of guaranteeing precise motion control by ensuring high confidence in the model during the operation.

Finally, the thesis analyses how design considerations affect the performance of surgical robotic systems in executing a desired operation. Generally, surgical robots are used in conjunction with serial-link manipulators, known as macro-micro manipulators. In this setup, the mounting connection between the two robots and the location of the insertion point highly affect the whole system’s performance. To this end, the thesis presents two optimization frameworks to find optimal configurations and ensure high accuracy, dexterity, and safety during the operation.

In fields like surgery where little or no sensors can be used, controllers must highly rely on the accuracy of the robot model. The novel insights shown in this thesis provide a step towards safer and more precise control of surgical robots, which is of paramount importance given the rising interest in increasing the level of autonomy in the operating room. Additionally, the strong push to reducing invasiveness is leading to the design of complex system like soft and continuum robots, whose modelling and control is even more challenging. We therefore believe that the approaches proposed in this thesis will be highly beneficial for enhancing the capabilities of the surgical robots of the future.