A Modular Synthetic Data Generation Toolkit for Vision-Based Localization Models in Intraocular Robotic Microsurgery

Intraocular microsurgery requires submillimeter precision within an extremely confined and delicate anatomical workspace. Cable-driven continuum robots such as Improved Integrated Robotic Intraocular Snake (I²RIS) offer the necessary dexterity but exhibit nonlinear hysteresis, complicating accurate control and localization. To enhance control precision, vision-based localization methods can be incorporated to provide external feedback and to compensate for modeling uncertainties. To support the development and quantitative evaluation of such vision-based approaches, modular, and open-source simulation framework is established, replicating an ophthalmic surgical scene that includes the eyeball model, the I²RIS continuum robot, and a calibrated surgical microscope. This environment enables automated and scalable acquisition of data that includes synchronized RGB-D images and corresponding ground-truth 6D pose data under diverse lighting, texture, and background conditions. Using the generated dataset, we conduct representative experiments on geometry-driven 6D pose tracking and appearance-based sim-to-real detection to evaluate its applicability for vision-based localization tasks. In addition, a preliminary domain gap analysis is performed using structure-based image similarity metrics, including Canny edge statistics and structural similarity (SSIM), to quantitatively assess visual alignment between simulated and real microscope images. The resulting dataset serves as a consistent evaluation resource for debugging, training, and assessing localization algorithms in intraocular continuum robotics, supporting reproducible research and sim-to-real generalization studies.