ETD PDF

Using Intelligent Anticipation to Improve Error-Prone Communication in Social Robots

Citation

Marulanda, Juan Felipe. (2020-05). Using Intelligent Anticipation to Improve Error-Prone Communication in Social Robots. Theses and Dissertations Collection, University of Idaho Library Digital Collections. https://www.lib.uidaho.edu/digital/etd/items/marulanda_idaho_0089e_11871.html

Title:
Using Intelligent Anticipation to Improve Error-Prone Communication in Social Robots
Author:
Marulanda, Juan Felipe
ORCID:
0000-0002-2473-0193
Date:
2020-05
Keywords:
Anticipation Communication Cooperation Navigation Robots Voice Recognition
Program:
Computer Science
Subject Category:
Artificial intelligence; Robotics; Engineering
Abstract:

This dissertation is centered on the study of anticipation behaviors based on Artificial Intelligence (AI) strategies as a method to improve the behavior of groups of autonomous ground robots by inferring the missing messages that get lost due to a noisy outdoor environment. These missing messages represent the information that the robots share as a group in order to stay synchronized. Also, this dissertation includes the study of an AI strategy as a method to recover corrupted phonetic messages. This research is inspired on the results of my Master thesis and expands on it to include more advanced anticipation techniques with a more complex navigation and sensors system for large environments; this research includes the use of GPS and compass for navigation and a (simulated) outdoor environment that is more realistic for many robotic applications and makes the inputs to the navigation system of the robot much noisier and more complex. This research also uses a small fleet of social robots whose collective behavior is less predictable and more complex. These issues make accurate communication between the robots a key and important feature and increase the need for anticipation as an aid to recover from faulty messages and events. A faulty event occurs when the communication between the robots gets interrupted and a message does not reach its destination.

In addition, this work expands the study of anticipation from two robots (one leader and a follower) to larger groups of robots (one leader and multiple followers). In this case, the followers attempt to stay in formation while the group navigates between waypoints. This means that this group of robots must coordinate their navigation control with their formation control in a complex, simulated outdoor environment. The anticipation models presented in this research are based on Fuzzy Logic and Artificial Neural Networks models (ANN). The last one was trained using methods like Backpropagation and a Genetic Algorithm. We also designed the anticipation models based on two structures that are commonly used in System Identification Theory: AutoRegressive with eXogeneous input (ARX) and AutoRegressive Moving Average with eXogenous input (ARMAX). This research includes tests of how well anticipation works to improve coordination in complex environments.

Also, we introduced an additional approach to message error correction based on syntax and phonetic inference as a complementary tool to message anticipation. This is based on the idea that some messages are not lost completely but their content can get corrupted. This approach uses an inference-based approach by implementing fuzzy logic theory in order to fully recover these corrupted messages. To test this idea, an approach based on human-robot interaction through voice commands was implemented. Here, external noise or even a user with a strong accent can confuse a speech engine and produce bad data. We define this as corrupted data and we use our solution to fix it.

The results showed that the ARMAX models were more successful in reducing the distance error between both a pair of robots and a larger group of robots than the ARX models while the Leader Robot run missions included in the training data. On the other hand, the ARX models were more successful than the ARMAX models in developing a generalization behavior while the Leader Robot run missions that were not similar to the training data. Also, we observed that the ANN models that were trained with Genetic Algorithms had better results than the ANN models trained with Backpropagation. We also noticed that the robots were able to recover their formation from collisions that happened during their path. But they still can only take a certain amount of collisions before running out of time to reorganize themselves. In addition, the results from the speech experiments showed that our approach was successful in recovering corrupted messages created by a speech engine after generating homophone words in a voice command.

Description:
doctoral, Ph.D., Computer Science -- University of Idaho - College of Graduate Studies, 2020-05
Major Professor:
Soule, Terence
Committee:
Edwards, Dean; Heckendorn, Robert; Zadehgol, Ata
Defense Date:
2020-05
Identifier:
Marulanda_idaho_0089E_11871
Type:
Text
Format Original:
PDF
Format:
application/pdf

Contact us about this record

Rights
Rights:
In Copyright - Educational Use Permitted. For more information, please contact University of Idaho Library Special Collections and Archives Department at libspec@uidaho.edu.
Standardized Rights:
http://rightsstatements.org/vocab/InC-EDU/1.0/