Online Learning of Industrial Manipulators' Dynamics Models

Research output: ResearchPh.D. thesis

Abstract

The robotics industry has introduced light-weight compliant manipulators to increase the safety during human-robot interaction. This characteristic is achieved by replacing the stiff actuators of the traditional robots with compliant ones which creates challenges in the analytical derivation of the dynamics models. Those mainly derive from physics-based methods and thus they are based on physical properties which are hard to be calculated. 
In this thesis, is presented, a novel online machine learning approach  which is able to model both inverse and forward dynamics models of industrial manipulators. The proposed method belongs to the class of deep learning and exploits the concepts of self-organization, recurrent neural networks and iterative multivariate Bayesian regression. It has been evaluated on multiple datasets captured from industrial robots while they were performing various tasks. Also, it was compared with multiple other state-of-the-art machine learning algorithms. Moreover, the thesis presents the application of the proposed learning method on robot control for achieving trajectory execution while learning the inverse dynamics models  on-the-fly . Also it is presented the application of the forward dynamics models on a model-based reinforcement learning scheme.
Close

Details

The robotics industry has introduced light-weight compliant manipulators to increase the safety during human-robot interaction. This characteristic is achieved by replacing the stiff actuators of the traditional robots with compliant ones which creates challenges in the analytical derivation of the dynamics models. Those mainly derive from physics-based methods and thus they are based on physical properties which are hard to be calculated. 
In this thesis, is presented, a novel online machine learning approach  which is able to model both inverse and forward dynamics models of industrial manipulators. The proposed method belongs to the class of deep learning and exploits the concepts of self-organization, recurrent neural networks and iterative multivariate Bayesian regression. It has been evaluated on multiple datasets captured from industrial robots while they were performing various tasks. Also, it was compared with multiple other state-of-the-art machine learning algorithms. Moreover, the thesis presents the application of the proposed learning method on robot control for achieving trajectory execution while learning the inverse dynamics models  on-the-fly . Also it is presented the application of the forward dynamics models on a model-based reinforcement learning scheme.
Original languageEnglish
Number of pages183
StatePublished - 2017
Publication categoryResearch

Bibliographical note

The author does not wish for his thesis to be published.

    Research areas

  • Robot Learning, Model Learning, Recurrent Neural Networks , Bayesian Inference, Reinforcement Learning, Industrial Robotics
ID: 261840199