Sim-to-real domain adaptation based completion level recognition for autonomous micro-drilling in biomedical application

Micron-level precision in drilling thin bone structures is essential in both surgical and neuroscience applications but remains technically challenging due to tissue fragility and anatomical variability. Traditional manual methods are slow and error-prone, while systems based on preoperative imaging lack adaptability to intraoperative changes. We previously proposed a convolutional neural network based autonomous micro-drilling system, enabling real-time control without prior bone knowledge. However, its reliance on manual annotation limited scalability and accuracy. In this study, we enhance the system through a sim-to-real domain adaptation model using synthetic data generated from a photorealistic simulator. The novelty of this work is a task-specific adversarial model that bridges the domain gap, significantly reducing annotation time (from 600 s/frame to 1.8 s/frame) while achieving a success rate of 85% (up from 80%) in 20 trials of eggshell drilling. These results demonstrate the feasibility and effectiveness of simulation-based training and domain adaptation for improving autonomous micro-drilling performance in biomedical applications.