| Soft robots are robots mainly made of materials with low Young’s modulus such as silicon.They are promising in applications such as medical care and domestic service,due to their advantages such as safety,compliance,and flexibility.However,due to their complex structures and non-linearity of their materials,controlling soft robots is still a challenge.Machine learning methods such as reinforcement methods have shown great potential in controlling soft robots.However,due to their poor sample efficiency,reinforcement learning methods require a large collection of training data,which may cost a lot of time to collect on a real-world soft robot.In this thesis,we proposed a novel Q-learning motion method for soft robots which can utilize a rough model to generate simulation data to train the Q-learning method to decrease the requirement of real-world data.Specifically,a rough model for a soft robotic arm and a Q-learning controller is proposed.Firstly a soft robotic model based on simplifying assumptions and Piece-wise constant curvature assumption is proposed,and experiments are performed to find out that the proposed model has a poor kinematics accuracy,but has a decent differential kinematics accuracy,so that it can be used for preliminary training of the controller.Besides,a series experiments are performed to analyse the influence of simplifying assumptions.Besides,in this thesis we also propose a Q-learning motion controller which learns the differential kinematics model of the robot.The controller firstly utilize the model for pre-training,and then utilize the real-world data collected during the control process to remedy the error of the model.A method to augment data is also proposed to further reduce the need of real-world training data.Finally,a series of experiments are performed to confirm that the controller can be used on a soft robotic arm with a lot of actuators,and to analyze the influence of pre-training,on-line learning,and data augmentation. |
PDF Full Text Request