Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)

First Advisor

Curtis Spenny, PhD


A simple edge-mating task, performed automatically by accommodation control, was used to study the feasibility of using data collected during a human demonstration to train an artificial neural network (ANN) to control a common robot manipulator to complete similar tasks. The 2-dimensional (planar) edge-mating task which aligns a peg normal to a fiat table served as the basis for the investigation. A simple multi-layered perceptron (MLP) ANN with a single hidden layer and linear output nodes was trained using the back-propagation algorithm with momentum. The inputs to the ANN were the planar components of the contact force between the peg and the table. The outputs from the ANN were the planar components of a commanded velocity. The controller was architected so the ANN could learn a configuration-independent solution by operating in the tool-frame coordinates. As a baseline of performance, a simple accommodation matrix capable of completing the edge- mating task was determined and implemented in simulation and on the PUMA manipulator. The accommodation matrix was also used to synthesize various forms of training data which were used to gain insights into the function and vulnerabilities of the proposed control scheme. Human demonstration data were collected using a gravity-compensated PUMA 562 manipulator and using a custom-built planar, low-impedance motion measurement system (PLIMMS).

AFIT Designator


DTIC Accession Number