Hierarchical Decision Framework Integrating Adaptive Task Planning with Trajectory Generation
- 주제(키워드) Reinforcement Learning , Imitation Learning , Large Language Model , Task Planning
- 주제(DDC) 004.6
- 발행기관 아주대학교 일반대학원
- 지도교수 Soyi Jung
- 발행년도 2026
- 학위수여년월 2026. 2
- 학위명 석사
- 학과 및 전공 일반대학원 AI융합네트워크학과
- 실제URI http://www.dcollection.net/handler/ajou/000000035780
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Autonomous excavation demands integrated decision-making spanning adaptive task planning, spatial execution, and trajectory generation. This thesis presents a hierarchical framework combining large language models, reinforcement learning, and attention-enhanced imitation learning for construction automation. The framework separates excavation control into two specialized layers operating at distinct decision-making scales. The strategic planning layer integrates LLM-based geometric decomposition with reinforcement learning, employing phase-guided action masking for adaptive target selection. The execution layer utilizes attention-enhanced imitation learning to produce adaptive trajectories across varying soil conditions. This architecture enables independent layer optimization while maintaining coordinated performance through structured information exchange. Validation in physics-based simulation achieves 93.3% success rate with 195.1mm spatial accuracy across diverse scenarios, demonstrating that specialized learning within hierarchical frameworks enables robust autonomous excavation for construction deployment.
more목차
1 Introduction 1
1.1 Background and Motivation 1
1.2 Contributions 3
2 Related Work 4
2.1 Trajectory Generation for Excavation Systems 4
2.2 Excavation Planning Approaches 5
2.3 Language Model-based Strategic Planning 6
3 Hierarchical Decision Framework for Autonomous Excavation 8
3.1 Overall Architecture 8
3.2 Hierarchical Layer Components 11
3.2.1 LLMaskRL: Strategic Planning Layer 11
3.2.2 AEIL: Execution Layer 12
4 LLMaskRL: Strategic Planning Layer 14
4.1 RL-based Spatial Planning 16
4.1.1 Proximal Policy Optimization 16
4.1.2 Planning Environment and State Representation 18
4.1.3 Action Space and Reward Structure 20
4.2 LLM-based Geometric Decomposition 22
4.2.1 Strategy Selection and Prompt Engineering 23
4.2.2 Phase-guided Action Masking 25
4.3 Integrated Execution Protocol 29
5 AEIL: Execution Layer 30
5.1 AEIL-based Trajectory Generation 32
5.1.1 Generative Adversarial Imitation Learning 33
5.1.2 Trajectory Generation Environment 34
5.1.3 State-Action Space Design 36
5.2 MHA-Enhanced Architecture Design 37
5.2.1 Multi-Head Attention Mechanism 38
5.2.2 Training Protocol and Data Augmentation 40
5.3 Integrated Execution Protocol 42
6 Performance Evaluation 45
6.1 Experimental Setup 45
6.2 Execution Layer Performance 49
6.2.1 Trajectory Generation across Multiple Scenarios 50
6.2.2 Comparative Analysis with Baseline Methods 54
6.2.3 Environmental Adaptability Validation 61
6.3 Strategic Planning Layer Performance 64
6.3.1 Geometric Decomposition Validation 65
6.3.2 Phase-guided Action Masking Effectiveness 66
6.3.3 Integrated System Performance Comparison 69
7 Physics-based Validation in Vortex Studio 73
7.1 Experimental Setup 73
7.2 Execution Layer Performance 74
7.2.1 Trajectory Generation across Multiple Scenarios 75
7.2.2 Comparative Analysis with Baseline Methods 77
7.3 Strategic Planning Layer Performance 81
8 Conclusion 88
References 90
