Author
Chen, Shuyang
Other Contributors
Wen, John T.; Diagne, Mamadou L.; Julius, Anak Agung; Mishra, Sandipan;
Date Issued
2021-05
Subject
Mechanical engineering
Degree
PhD;
Terms of Use
This electronic version is a licensed copy owned by Rensselaer Polytechnic Institute, Troy, NY. Copyright of original work retained by author.;
Abstract
Typically for industrial robots, users have no access to motor torque control. Instead, a robotmay be controlled by an external position or velocity command interface. This outer-loop external
interface facilitates sensor-guided and robotic shared motion control. However, the robot trajectory
tracking performance is affected by the unknown robot dynamics and joint servo controllers.
The performance is further compromised by time delays in transmitting the external command as a
setpoint to the inner control loop. Thus, it is necessary to compensate for the internal dynamics to
improve the motion control performance. However, it is in general a difficult problem to analytically
identify a highly nonlinear and coupled robot dynamical model. Instead, a universal function
approximator such as a neural network may be adopted for modeling the system dynamics.; Robot manipulators have been widely utilized in industrial automation to improve productquality and productivity. Common robotic tasks in an industrial setting include assembly, welding,
surface treatment, packaging and labeling. The various tasks also stimulate the development of
sophisticated control algorithms and dexterous tools for robot manipulators.; This thesis investigates the application of optimization and deep learning to robotics motionand force control. First, by leveraging a position control interface, a kinematic controller is developed
by solving a quadratic-programming problem. However, the motion control performance is
obviously compromised by the unknown internal dynamics when the robot is commanded a fast
trajectory. To improve the motion control accuracy, we then develop a model-free gradient-based
iterative learning control (ILC) algorithm with the gradient descent directions generated by the
physical system. However, the algorithm may be time-consuming considering iterative updates,
and the learned representation is not transferable to a new trajectory. This motivates the use of
neural networks to approximate the system dynamics to improve the time efficiency and generalization
of ILC. And a neural-learning control framework is proposed to address robot motion
control in dynamics level through dynamics compensation using neural networks. Finally, the
control scenario is extended from robot motion control to hybrid force and motion control, motivated
by a robotic deep rolling task. We apply the model-free ILC algorithm to deep rolling on flat
and curved surfaces.;
Description
May 2021; School of Engineering
Department
Dept. of Mechanical, Aerospace, and Nuclear Engineering;
Publisher
Rensselaer Polytechnic Institute, Troy, NY
Relationships
Rensselaer Theses and Dissertations Online Collection;
Access
Users may download and share copies with attribution in accordance with a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 License. No commercial use or derivatives are permitted without the explicit approval of the author.;