Motion Primitives Segmentation for A Tendon-driven Continuum Manipulator through Imitation Learning Process
Continuum robots have been widely developed to enhance the performance of robot-assisted surgery and minimally invasive surgery (MIS). They are always made of soft compliance materials with flexible actuation mechanisms and accessible to the confined surgical cavity. Meanwhile, they are introduced with inherent nonlinearity and uncertainties. Therefore, accurate kinematics models and motion planning for continuum robots become extremely challenging. To deal with the dilemma, machine learning algorithms have been introduced for solving continuum robot’s kinematics, the mapping between the actuation space, task space, and configuration space. Imitation learning is such a data-driven approach. Users provide demonstrations to show a specific motion, then the consistency of the several demonstrations would be encoded into a statistical model and the meaningful information would be extracted out. Imitation learning directly aims at the search space for policy learning, which enables us to understand the mechanisms between motor action and task performance. Imitation learning is a powerful method to encode inverse kinematics of continuum manipulators.
Moreover, hybrid utilization of machine learning techniques is required when planning motion for continuum robots. It is reasonable to segment the motion data into various motion primitives and then analyze them. Classification is introduced to work out motion primitives, then they are developed with dynamical policies. To achieve high-dimensional data classification, Support Vector Machine (SVM) is considered. Since the classification process leaves out the temporal information, another model for sequencing the primitives is required. Hidden Markov Model (HMM) is proposed for achieving the transfer among a number of motion primitives based on probability derived from samples. Finally, a hybrid machine learning model will build for the continuum robot’s motion planning.
Our experiment of feedforward backpropagation (FF-BP) neural network for predicting a tendon-driven continuum manipulator (TCM) kinematics has shown that machine learning with potential to encode kinematics for continuum manipulators. Then we plan to work out a motion primitive library as well as a hybrid machine learning model for the TCM. To verify the applicability of hybrid SVM/HMM model, we will execute a series of experiments. In conclusion, the research work focuses on developing an accurate forward kinematics as well as inverse kinematics for continuum manipulators and building diverse motion primitives for continuum manipulator’s motion planning. The significance of this research work is to improve the autonomy of the continuum manipulators, to develop its potential application in minimally invasive surgery.
Miss WANG Jiao
9:00am – 11:00am