Optical active perception with 3D Gaussian splatting enables autonomous instrument insertion in robotic vitreoretinal surgery

We present an optical active perception framework that enables autonomous instrument insertion through a trocar using robot-guided multi-view imaging. A low-cost RGB-D camera provides depth-guided viewpoint planning, while 3D Gaussian Splatting (3DGS) reconstructs a high-resolution trocar model from sparse multi-view RGB images with submillimeter fidelity, overcoming the practical limitations of commercial 3D optical sensing systems, such as stereo and structured-light cameras, in this constrained surgical scenario. Integrated with a 6-DOF robotic arm, the system estimates the trocar pose and aligns the instrument for insertion. Experiments on static and motion-simulated ex vivo porcine eyes achieve consistent submillimeter positional accuracy and angular accuracy on the order of a few degrees, sufficient for safe insertion given the narrow trocar clearance. These results establish the potential of 3DGS-based optical active perception for robotic microsurgical navigation.