Continuous Control of a Soft Continuum Arm using Deep Reinforcement Learning

Sreeshankar Satheeshbabu, Naveen K. Uppalapati, Tianshi Fu, Girish Krishnan

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Soft Continuum Arms (SCAs) are challenging to control due to their highly nonlinear characteristics and sensitivity to external loading. Recent efforts to address the control problem using machine learning techniques are limited to simple SCA architectures. In this paper, we train a model-free reinforcement learning control policy based on Deep Deterministic Policy Gradient (DDPG) for end effector path tracking on a BR2 SCA. Unlike simple SCA architectures, the BR2 SCA has the functionality to bend and rotate spatially thus leading to enhanced workspace and ability to perform complex tasks. The control policy is first validated in simulations and then implemented on a prototype BR2 with state feedback. An average tracking error less than 3 cm (< diameter of the SCA) is reported using the proposed control policy. The efficacy of the control policy is validated for different loading conditions both in simulations and on the SCA prototype.

Original languageEnglish (US)
Title of host publication2020 3rd IEEE International Conference on Soft Robotics, RoboSoft 2020
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages497-503
Number of pages7
ISBN (Electronic)9781728165707
DOIs
StatePublished - May 2020
Event3rd IEEE International Conference on Soft Robotics, RoboSoft 2020 - New Haven, United States
Duration: May 15 2020Jul 15 2020

Publication series

Name2020 3rd IEEE International Conference on Soft Robotics, RoboSoft 2020

Conference

Conference3rd IEEE International Conference on Soft Robotics, RoboSoft 2020
CountryUnited States
CityNew Haven
Period5/15/207/15/20

ASJC Scopus subject areas

  • Artificial Intelligence
  • Mechanical Engineering
  • Control and Optimization

Fingerprint Dive into the research topics of 'Continuous Control of a Soft Continuum Arm using Deep Reinforcement Learning'. Together they form a unique fingerprint.

Cite this