Title Page
Abstract
Contents
Chapter 1. Introduction 11
1.1. Background and Motivation 11
1.2. Objective and Contribution 16
1.3. Organization 17
Chapter 2. Related Work 18
2.1. Localization 18
2.2. Wheel Slippage Estimation 22
Chapter 3. Analysis of Lugged-wheel Motion 24
3.1. Dynamic model of lugged-wheel 27
3.2. Simulation of lugged-wheel motion 33
Chapter 4. Proposed Localization Method 38
4.1. Overall structure 38
4.2. Structure of IEKF 40
4.3. DNN-based slip ratio estimator 45
Chapter 5. Experiments 51
5.1. Experimental setup 52
5.2. Validation of simulation 63
5.3. Training methodology 67
5.4. Validation results 71
5.5. Experimental results 79
5.6. Computing time 86
5.7. Future work 87
Chapter 6. Conclusion 89
Appendices 91
Appendix A. The Absence of Inertial Sensor Biases in the State of IEKF 91
Appendix B. Estimating Slip Ratio Between Classification and Regreesion 97
Appendix C. Performance Evaluation of a Smoothing Filter Usage 103
Appendix D. Performance Evaluation of a Moving Average Filter 110
Appendix E. Target Data of DNN-Based Slip Ratio Estimator 113
Appendix F. Localization Result of Test Dataset 115
Appendix G. Heading Error of Test Dataset 117
Bibliography 118
초록 132
Table 1. Detail of the CNN 50
Table 2. Detail of the FC layer 50
Table 3. The detail of training datasets. 69
Table 4. The detail of the scenario. 69
Table 5. Validation results. 74
Table 6. Experimental results. 81
Table 7. Computing time. 86
Table A.1. Standard deviation of inertial sensor for stationary condition 95
Table A.2. Standard deviation of inertial sensor for the scenario involving 1 minute of stationary condition, followed by 40 seconds of... 95
Table A.3. Average inertial sensor data for the scenario involving 1 minute of stationary condition, followed by 40 seconds of movement, and then... 96
Table B.1. Detail of the FC layer of FC layer-based slip ratio estimator 101
Table C.1. Experimental results dependent on usage of smoothing filter. 109
Table C.2. Computing time. 109
Table D.1. Experimental results dependent on usage of moving average filter. 111
Table F.1. Experimental results for all scenario. 116
Table G.1. Experimental results of heading errors. 117
Figure 1.1. Various types of terrestrial robot. The types are classified according to various form of terrestrial locomotion, which are walking,... 12
Figure 1.2. Wheel slippage experience by the lugged-wheel robot. 13
Figure 1.3. Various sensors used for localization. 15
Figure 2.1. A schematic depiction of optimization-based method and Kalman filter-based method. 20
Figure 2.2. Robot platform for wheel slippage experiments. 23
Figure 3.1. Ideal and non-ideal displacement in wheel deformation and terrain deformation. The displacement difference due to slippage is... 26
Figure 3.2. The model and variable of the lugged-wheel motion. The graphical representation of coordinates, x and θ are shown. 28
Figure 3.3. The force and torque acting on the lugged-wheel and robot for (a): case 1 (no-slip case), 2 (slip case) and (b): case 3 (two-lugs case). 29
Figure 3.4. Velocity of the robot and the lugged-wheel for simulation. (a) Overall simulation results. (b) Simulation result magnified to display only... 35
Figure 4.1. Overall structure of proposed method. 39
Figure 4.2. Coordinate system where OXYZ represents the world frame and oxyz represents the lugged-wheel robot frame. 40
Figure 4.3. Overall structure of DNN-based slip ratio estimator. 46
Figure 5.1. Lugged-wheel robots and lugged-wheel. The robot platforms are shown in (a) and (b). The lugged-wheel of robot platform is shown in (c). 53
Figure 5.2. The satellite view of the outdoor experiments for training datasets, along with the locations corresponding to each slope. 55
Figure 5.3. Outdoor environments for the slope 1. 56
Figure 5.4. Outdoor environments for the slope 2. 57
Figure 5.5. Outdoor environments for the slope 3. 58
Figure 5.6. Outdoor environments for the slope 4. 59
Figure 5.7. The satellite view of the outdoor experiments for testing datasets, along with the locations corresponding to each slope. 60
Figure 5.8. Outdoor environments for the slope 5 and 6. 61
Figure 5.9. Experimental result for velocity of the wheel and robot on tile. 64
Figure 5.10. Experimental result for velocity of the wheel and robot on (a) tile and (b) carpet. 65
Figure 5.11. Experimental result for velocity of the wheel and robot on carpet. 66
Figure 5.12. Example of slip ratio of IEKF-based localization and customized slip ratio 70
Figure 5.13. Slip ratio of IEKF-based localization, DNN-based slip ratio estimator, and customized slip ratio for (a) dataset no. 2-1 and (b) dataset no. 2-5. 75
Figure 5.14. Slip ratio of IEKF-based localization, DNN-based slip ratio estimator, and customized slip ratio for (a) dataset no. 2-8 and (b) dataset no. 2-13. 76
Figure 5.15. Trajectories of integration-based localization, IEKF-based localization, and proposed method for dataset no. 2-13. 77
Figure 5.16. Slip ratio of IEKF-based localization for dataset no. 2-13 and 3-13. 78
Figure 5.17. Trajectories of integration-based localization, IEKF-based localization, and proposed method for (a) dataset no. 4-1 and (b) dataset no. 4-3. 82
Figure 5.18. Trajectories of integration-based localization, IEKF-based localization, and proposed method for (a) dataset no. 4-5 and (b) dataset no. 4-7. 83
Figure 5.19. Trajectories of integration-based localization, IEKF-based localization, and proposed method for (a) dataset no. 4-9 and (b) dataset no. 4-12. 84
Figure 5.20. Example of slip ratio of IEKF-based localization and DNN-based slip ratio estimator for dataset no. 4-9. 85
Figure A.1. (a) acceleration and (b) angular velocity of the inertial sensor for stationary condition. 93
Figure A.2. (a) acceleration and (b) angular velocity of the inertial sensor for the scenario involving 1 minute of stationary condition, followed by 40... 94
Figure B.1. Overall structure of DNN-based slip ratio estimator (Regression version). 100
Figure B.2. Overall structure of FC layer-based slip ratio estimator (Regression version). 101
Figure B.3. Slip ratio for regression approach. (a) Regression version DNN- based slip ratio estimator. (b) FC layer-based slip ratio estimator. 102
Figure C.1. Slip ratio for Slope no. 1. (a) Result of slip ratio estimator not including smoothing filter. (b) Result of slip ratio estimator including... 105
Figure C.2. Slip ratio for Slope no. 2. (a) Result of slip ratio estimator not including smoothing filter. (b) Result of slip ratio estimator including... 106
Figure C.3. Slip ratio for Slope no. 3. (a) Result of slip ratio estimator not including smoothing filter. (b) Result of slip ratio estimator including... 107
Figure C.4. Slip ratio for Slope no. 4. (a) Result of slip ratio estimator not including smoothing filter. (b) Result of slip ratio estimator including... 108
Figure D.1. Calculated slip ratio and slip ratio after using moving average. 112