본문 바로가기 주메뉴 바로가기
22대 대한민국 국회 국회정보길라잡이 국회의원검색
회원가입 로그인
HOME

정보검색

소장정보 검색

국회도서관 홈으로 정보검색 소장정보 검색

결과 내 검색

동의어 포함

고급검색

인명/단체명

저자정보 상세정보
인명/단체명을 입력하세요.

키워드

  • 대표어

  • 외국어

-

목차보기

Title Page 2

Abstract 6

Contents 7

Chapter 1. Introduction 12

1.1. Insect-inspired Robots 12

1.1.1. Research Trends 14

1.1.2. Challenges of Intelligent Control in Small-scale Robots 16

Chapter 2. Methodology 18

2.1. Bio-inspired Mechanosensory System 18

2.1.1. Campaniform Sensilla 18

2.1.2. Ultrasensitive Crack-based Strain Sensor 20

2.2. Deep Learning Model and Data Processing 22

2.2.1. 1D Convolutional Neural Network (CNN) 23

2.2.2. Exploratory Data Analysis (EDA) 24

2.3. End-to-End Reinforcement Learning 26

2.3.1. Soft Actor-Critic (SAC) Algorithm 27

2.3.2. Design of Reward Function 28

Chapter 3. Results 32

3.1. Control of Flapping-wing Drone 32

3.1.1. Wind Vector Prediction 33

3.1.2. Control in a 1 DOF Movement Environment 36

3.1.3. Control in a 2 DOF Movement Environment 39

3.1.4. Control in a 6 DOF Movement Environment 42

3.1.5. Control in a Windless Environment 46

3.2. Control of Legged Robot 50

3.2.1. Experimental Setup 52

3.2.2. Terrain Type Classification 55

3.2.3. Robot Orientation Regression 56

3.2.4. Energy-efficient Control Across Various Terrains 59

3.2.5. Adaptive Control on Mixed Terrain 62

Chapter 4. Conclusion and Future Work 66

4.1. Summary 66

4.2. Future Work 67

Bibliography 70

List of Tables 11

Table 2.1. Comparison of reinforcement learning algorithm for the system 28

Table 3.1. Mechanical properties of each terrain 53

Table 3.2. Comparison of power saving across tested terrains 65

List of Figures 9

Figure 2.1. Three main components of strain-based intelligent control system in this study 18

Figure 2.2. Stress distribution in a simplified cantilever beam model, indicating that campaniform... 19

Figure 2.3. Attachment of the bio-inspired strain sensor to the wing base and leg joint of the robot 20

Figure 2.4. Fabrication process of the crack-based sensor and its compact form factor suitable for... 21

Figure 2.5. Higher signal-to-noise ratio than a commercial strain gauge 22

Figure 2.6. Sensor signals exhibiting nonlinear responses to external environmental changes... 23

Figure 2.7. 1D CNN architecture for feature extraction from time series nonlinear data 24

Figure 2.8. Strain signal reflection of the subtle change of wind vector from front, right, and left 25

Figure 2.9. Gaussian distributed motor power to measure diverse condition of sensor signals 26

Figure 2.10. Reward function strategy of left, right, circular, and s-curve path control 29

Figure 2.11. Reward function strategy of lower, and higher altitude path control 30

Figure 2.12. Reward designation from the cost of transport (CoT) 31

Figure 3.1. Overview of strain-based intelligent control for flapping-wing drone 33

Figure 3.2. 186 cases of wind vectors to predict from the wing strain signals 34

Figure 3.3. Result of regression and classification of the wind vectors 35

Figure 3.4. Training architecture for 1 DOF movement control 36

Figure 3.5. Converging trends of target angle and score as the number training episode increases 38

Figure 3.6. Control strategy to sustain the target position in 1 DOF movement environment 39

Figure 3.7. Training architecture for 2 DOF movement control 40

Figure 3.8. Recovering ability of trained model in the falling state 41

Figure 3.9. Converging trends of score, rotational and pitch angle of the drone as the number... 42

Figure 3.10. Training strategy of 6 DOF movement control 43

Figure 3.11. Wind tunnel setup of turbulence and reward function for target position control 43

Figure 3.12. Comparison of visited points at each step between untrained model and trained models 44

Figure 3.13. Control strategy of the trained model flying along the left-side route to reach the target... 45

Figure 3.14. Control ability of trained model when change the target position to left or right side 46

Figure 3.15. Path control strategy of flapping-wing drone in a windless environment 47

Figure 3.16. Take-off shooter that initializes when each episode begins 48

Figure 3.17. Path control result of s-curve and circular motion and the predicted odometry 49

Figure 3.18. Preprocessing of deep learning dataset for odometry prediction 50

Figure 3.19. Overview of strain-based intelligent control for a legged robot 51

Figure 3.20. Experimental setup of the legged robot for training and control 52

Figure 3.21. Customized legged robot with strain sensor and motion capture-based reward... 54

Figure 3.22. Control diagram of the legged robot according to the reinforcement learning 55

Figure 3.23. Confusion matrix result of the terrain type classification 56

Figure 3.24. Strain signal input used for pitch and roll angle prediction during running 57

Figure 3.25. Prediction result of roll and pitch angle errors 58

Figure 3.26. Prediction result of roll and angle errors expressed as percentages 59

Figure 3.27. Converging trend of average CoT across terrains with increasing training episodes 61

Figure 3.28. Box plot comparison of CoT between untrained and trained models 62

Figure 3.29. Converging trend of average CoT and box plot comparison in mixed terrain 63

Figure 3.30. Adaptable control performance in the terrain transition region for efficient running 64

Figure 4.1. Preservation of prediction output after quantization of the optimized policy network 68

Figure 4.2. Demonstration of the feasibility of running a trained policy network on a memory-... 69

초록보기

 Insect-inspired robots are promising for tasks requiring compactness, agility, and multifunctionality, taking cues from organisms like flies and ants. Using lightweight frames, origami structures, and soft actuators such as shape memory alloys or dielectric elastomers, these robots can walk, jump, or fly efficiently at millimeter scales-ideal for confined or hazardous environments. However, miniaturization limits sensing and control: IMUs suffer from noise under strong vibrations, and vision systems degrade under occlusion, dust, or low light. To overcome these limitations, we propose a bio-inspired control framework using strain sensing, modeled after campaniform sensilla in insect wings and legs. Our method integrates three components: ultrasensitive crack-based strain sensors at mechanically critical joints such as wing bases, leg hinges, data-driven state estimation using 1D convolutional neural network (1D CNN), and Soft Actor-Critic (SAC) reinforcement learning (RL) algorithm for real-time control using only strain feedback. In aerial applications, a flapping­wing drone with just two wing-base strain sensors estimated wind direction and speed with approximately 80% classification accuracy and 29° mean angular error. It performed flight control in windy and windless environments-all without IMUs or cameras-by adapting to subtle changes in airflows solely based on strain signals. On the ground, a legged robot classified terrain with 93% accuracy and estimated pitch and roll with 〈4% error. An RL controller trained on this input minimized energy consumption (cost of transport, CoT) by up to 7.8% on low-friction terrain and adapted effectively to mixed surfaces. These results highlight the potential of strain-based proprioception as a lightweight, scalable alternative to conventional sensors. The unified approach enables adaptive, efficient control in both aerial and terrestrial microrobots operating in unstructured environments, paving the way for future autonomous systems with minimal sensor configurations.

추천서가 (다양한 추천 자료를 만나보세요)

권호기사보기

권호기사 목록 테이블로 기사명, 저자명, 페이지, 원문, 기사목차 순으로 되어있습니다.
기사명 저자명 페이지 원문 기사목차

권호

기사명 원문보기 기사목차
닫기