An Investigation of Predictive and Adaptive Model-Based Methods

for Direct Ground-to-Space Teleoperation with Time Delay

A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science By JEFFREY BRENT ELLIS B.S., University of Cincinnati, 1988 1998 Wright State University COPYRIGHT (C) BY JEFFREY B. ELLIS 1998

ABSTRACT

Ground-to-space teleoperation is hindered by time delays imposed by the speed-of-light limit of radio transmissions over large distances, as well as communication processing delays. Consequently, the human operator must wait to observe the results of each small movement before committing to further action. A predictor display, which presents the user with a simulation of the system rather than the time-delayed video/ telemetry data, can lengthen each step of a move-and-wait approach. Due to the accumulation of modeling errors, however, the predictor display does not negate the need for moving-and-waiting. The user must now wait until the simulation becomes resynchronized with the system before transmitting further commands.

This thesis explores the use of predictor displays to cope with time-delayed teleoperation. The thesis project consists of two phases: (1) the development of a testbed simulation for evaluating predictor display concepts; and (2) an investigation into the feasibility of improving the baseline predictor display's performance. Under Phase 1, a simulation of a teleoperated spacecraft was developed and validated, and the behavior of the baseline predictor method was characterized in detail. For the selected task, it was shown that improvements to the baseline predictor are required in order for continuous control to be realized. Under Phase 2, a Cerebellar Model Articulation Controller (CMAC) neural network was employed in an attempt to "learn" the error between the predictor and the actual system, so that the predictor display could be corrected accordingly. The results indicate that the error functions in both position and acceleration contain abrupt discontinuities that cannot be accurately modeled by the CMAC due to its local generalization behavior. These findings have led to the conclusion that other error modeling techniques are required, and/or that smoothing or filtering methods in the state domain must be implemented. As a result of the Phase 2 investigation, three proposed techniques that merit future investigation were identified:

Use multiple neural networks to predict the discontinuities in the error function and to model the piecewise continuous portions of the error function separately.
Use a mathematical constraint such as a spline function to force the predicted state along a continuous path.
Run the predictive model "open loop" and attempt to use subtle corrective commands to resynchronize the predictor and the system.

Acknowledgements

    1. Introduction
            1.1 Project Overview
                    1.1.1 Problem Statement
                    1.1.2 Teleoperation Simulator Development
                    1.1.3 Predictor Improvement Investigation
                    1.1.4 Summary of Results
            1.2 Motivation
            1.3 Background

    3. Teleoperation Experiment Implementation
            3.1 Design Considerations
                    3.1.1 Simulation Environment
                    3.1.2 Implementation Constraints
            3.2 Classes
                    3.2.1 Ground Segment Classes
                    3.2.2 Space Segment Classes
                    3.2.3 "Shared" Classes
            3.3 Program Operation
            3.4 Program Validation

    4. Candidate Predictor Improvements
            4.1 Overview of Adaptive Methods
            4.2 Adaptive Model Methods
                    4.2.1 On-Line Parameter Estimation
                    4.2.2 On-Line System Identification
            4.3 Adaptive Control Methods
                    4.3.1 Model-Reference Adaptive System
                    4.3.2 Self-Tuning Regulator
            4.4 Selection of Adaptive Methods

    6. Discussion and Conclusions
           6.1 Baseline Predictor Performance
           6.2 Error Estimation Methods
           6.3 Recommendations

Appendix A (not available on-line)

References

List of Figures

1.1. Task performance model

2.1.    Spacecraft schematic
2.2.    Spacecraft bus component layout
2.3.    Generalized teleoperation architecture
2.4.    Teleoperation architecture for acceleration command mode
2.5.    Teleoperation architecture for rate command mode
2.6.    Top-level model entity hierarchy
2.7.    Ground segment architecture
2.8.    Model entity interactive relationships
2.9.    Thruster selection example
2.10. Spacecraft model entities and attributes
2.11. Model entity interactive relationships
2.12. Exaggerated thrust vs. time history

3.1. Class hierarchy

4.1.    A neural network with 3 inputs and 2 outputs
4.2.    Conceptual structure of a CMAC
4.3.    Receptors for a single input dimension
4.4.    Layering of the address space A
4.5.    Conceptual architecture for the MRAS
4.6.    Conceptual architecture for the STR

5.1.    Discontinuity in predicted state vs. time
5.2.    Use of an error model to correct the predictor
5.3.    CMAC error model
5.4.    Angular acceleration vs. time
5.5.    Velocity vs. time
5.6.    Angle vs. time
5.7.    Angle error vs. time
5.8.    CMAC error estimation
5.9.    Actual and predicted acceleration
5.10. Actual and predicted velocity
5.11. Predicted and residual acceleration error

6.1. Smoothing the discontinuity

Back to my home page

An Investigation of Predictive and Adaptive Model-Based Methods

for Direct Ground-to-Space Teleoperation with Time Delay

Table of Contents

List of Figures