An imitation learning-based approach to the motion planning of robots : from discrete to continuum robots

Abstract

Robots have impacted human everyday life deeply over the past decades, from traditional automated manufacturing to the state-of-the-art robot-assisted surgery. With the advancement of robotic technologies, numerous different kinds of robots have been invented to deal with the increasingly complicated requirement. For instance, continuum robots have been developed to enhance the performance of robot-assisted minimally invasive surgery (MIS), which is one of the major focuses of this thesis. However, effective and efficient motion control of robots, including inverse kinematics modelling and motion planning, remains challenging, especially for the continuum robots. In this thesis, we propose an imitation learning-based framework to solve this issue, both discrete and continuum robots have been investigated. For discrete manipulators, conventional Jacobian-based methods are adopted to address the corresponding inverse kinematic problems. While for the continuum robot, we implement three machine learning algorithms, K-Nearest Neighbors Regression (KNNR), Gaussian Mixture Regression (GMR), and Extreme Learning Machine Regression (ELMR), to approximate its inverse kinematics model. We also evaluate the possibility of applying reinforcement learning to further improve the inverse kinematics model of continuum robots. Two different types of imitation learning-based frameworks have been investigated to plan the motion path of robots, namely model-free and model-based methods. In model-free imitation learning, human demonstrations are provided in both actuation space and task space, and Gaussian Mixture Model (GMM) and GMR are used to encode the demonstrations directly and then generalize an executable path for the robot to reproduce the learned task. For complicated tasks, the demonstration has to be segmented into multiple movement primitives to simplify the learning process, and Support Vector Machine (SVM) is used for the segmentation purpose. Dynamical systems have been deployed to facilitate the model-based method, whose effectiveness is ensured by Lyapunov Stability Theorem, and the demonstrations are only available in the task space. A number of robots have been developed in this thesis to evaluate the proposed approaches, including a re-designed 7-DoF Mitsubishi PA-10 robot, a 6-DoF ionic polymer-metal composite (IPMC) flexible manipulator, a 3-DoF tendon-driven continuum manipulator (TCM), and a 2-DoF TCM. Inverse kinematics of the 3-DoF TCM is learned by KNNR, GMR, and ELMR. While reinforcement learning has been applied to the 2-DoF TCM, and in a few iterations the trajectory tracking error reduces significantly from 8.045 mm to 1.101 mm. With model-free imitation learning, the 6-DoF IPMC manipulator acquires the skill of navigation through a narrow hole, and the 3-DoF TCM learns to reproduce two surgical related tasks, namely complaint tube insertion and simplified endoscopic submucosal dissection (ESD). While with model-based imitation learning, the PA-10 robot successfully bats a fast flying ball, a 7-DoF KUKA LBR iiwa robot learns to perform manipulation tasks under both spatial and temporal perturbations, and the 3-DoF TCM obtains the skill of obstacle avoidance. The significance of this thesis is three-fold. Firstly, machine learning and reinforcement learning are investigated to address the inverse kinematics of continuum robots. Secondly, motion planning of both discrete and continuum robots is facilitated via imitation learning. Lastly, both model-free and model-based imitation learning have been developed.