Deep reinforcement learning control of a 2D soft robotic arm

This thesis provides a deep reinforcement learning (DRL) based approach for the development of a control policy for a 2D soft robotic arm. The simulation is based on the SOFA framework, which is a real-time multi-physics simulation package capable of creating models and computing forces for deformable materials. The 2D soft robotic arm is composed of two modules where each module consists of two pneumatic actuators and can extend and bend. Though DRL has been explored in the soft robotics realm, end-to-end training has not been developed. Herein, this thesis presents an end-to-end training from snapshots of the simulation to control policy guiding the soft robotic arm to reach a designated target using DRL, and proofs the validity and stability of this approach. The soft robotic arm is able to reach the target with a 98.1% success rate after approximately 30 epochs of training both for fixed initial position training and varying initial position training. The methodology presented here can be generalized for intelligent motion planning and control of soft robotic systems that are otherwise challenging.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2020

Author dc:creator

• Shen, Zhongyi

Contributors dc:contributor

• Zhang, Yang

• deep Q network
• deep reinforcement learning
• soft robotic arm
• control

Statement dc:rights

• Copyright 2020 Zhongyi Shen

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/108623

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/108623