Iyengar, Keshav; (2023) Deep Reinforcement Learning for Inverse Kinematics and Path Following for Concentric Tube Robots. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text PhD_Thesis__Final_Version.pdf - Accepted Version Download (45MB) | Preview

Abstract

Concentric tube robots (CTRs) are continuum robots that allow for bending and twisting motions unattainable by traditional rigid link robots. The curvilinear backbones can benefit surgical applications by improving dexterity, enlarging the workspace, and reducing trauma at the entry point of the instrument. The curvilinear backbone that is attributed is the result of pre-curved, super-elastic tubes arranged concentrically. Each tube has a straight and pre-curved section and is actuated in rotation and translation from the tube base with the neighboring tube interactions producing the curvilinear backbone. The modeling of the neighboring tube interactions is non-trivial, and an explored topic in CTR literature. However, model-based kinematics and control can be inaccurate due to inherent manufacturing errors of the tubes, permanent deformation over time, and unmodelled physical interactions. This thesis proposes a model-free control method using deep reinforcement learning (DRL). The DRL framework aims to control the end-effector of the CTR with limited modeling information by leveraging simulation data, which is much less costly than hardware data. To develop a DRL framework, a Markov Decision Process (MDP) with states, actions, and rewards needs to be defined for the inverse kinematics task. First, action exploration was investigated with this MDP in a simpler simulation as CTRs have a unique extension degree of freedom per tube. Next, state representation, curriculum reward, and adaptation methods over multiple CTR systems were developed in a more accurate simulation. To validate the work in simulation, a noise-induced simulation environment was utilized to demonstrate the initial robustness of the learned policy. Finally, a hardware system was developed where a workspace characterization was performed to determine simulation to hardware differences. By using Sim2Real domain transfer, a simulation policy was successfully transferred to hardware for inverse kinematics and path following, validating the approach.

Download activity - last month

Export as XMLCSVJSON

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as JSONCSVXML

Archive Staff Only

View Item