Title

Markerless Eye Trocar Alignment Using a Vision-Language-Action Policy for Robot-Assisted Vitreoretinal Surgery

Alternative Title

로봇 보조 유리체망막 수술을 위한 VLA 정책을 활용한 마커리스 안구 트로카 정렬

DGIST Authors

Jongyeol Na ; Cheol Song

Advisor

송철

Issued Date

2026

Awarded Date

2026-08-01

Type

Thesis

Description

Autonomous eye trocar alignment, Vision-Language-Action（VLA）, Fiber-optic sensor, Markerless, Vitreoretinal surgery

Table Of Contents

1 Introduction 1
1.1 Study Background 1
1.2 Previous Studies 3
1.3 Study Objective 6
1.4 Study Scope 7
2 Working Principle 8
2.1 Trocar Docking Pipeline 8
2.2 Autonomous Demonstration Generation 10
2.3 Model Architecture and Loss Functions 12
2.3.1 Model Architecture 12
2.3.2 Loss Functions 13
3 System Design and Implementation 16
3.1 Hardware Architecture and System Setup 16
3.2 Preliminary Test 18
3.3 Digital Twin Pipeline and Data Generation 20
3.3.1 Random Phantom Position 22
3.3.2 Random Robot Initial Position 22
3.4 Implementation Details 23
4 System Evaluation and Results 25
4.1 Experimental Setup and Evaluation Protocol 25
4.2 Micro-alignment: Reaching versus Sustained Holding 27
4.3 Ablation Studies 30
4.3.1 Vision Encoder 30
4.3.2 Near-Goal Fine-Tuning 31
4.3.3 Data Composition and Workspace Robustness 32
4.3.4 Execution Horizon 32
4.3.5 Input Representation: Cropping and View Configuration 33
4.3.6 Auxiliary Losses and Trajectory Regularization 34
5 Conclusion 36
5.1 Summary 36
5.2 Limitations ＆ Discussion 37
5.3 Future Works 38
Appendix 39
A.1 Macro-Approach: Global Workspace Navigation 39
A.2 Frequency-Domain Analysis of the Action Trajectories 40
REFERENCES 42

URI

https://scholar.dgist.ac.kr/handle/20.500.11750/60822
http://dgist.dcollection.net/common/orgView/200001010815

DOI

10.22677/THESIS.200001010815

Degree

Master

Department

Artificial Intelligence Major

Publisher

DGIST

Show Full Item Record