Digital Twin-Driven OCT Probe Guidance via Group-Relative Soft Actor-Critic

To address the precision and safety demands of OCT probe guidance in subretinal injection surgery, we present a digital twin-based robotic visual servoing framework. By deploying deep reinforcement learning within a virtual mapping space, risks associated with physical trial-and-error interactions are effectively mitigated. To overcome policy convergence issues arising from sparse rewards, we propose the Group-Relative Soft Actor-Critic (GRSAC) algorithm. This method utilizes an intra-group relative ranking mechanism to extract robust, scale-invariant dominance signals. Integrated with an eye-to-hand visual synchronization interface, the framework facilitates seamless sim-to-real policy transfer. Experiments demonstrate that the system achieves preoperative coarse positioning with an average error of 0.81 cm in unstructured environments, validating the efficacy of the digital twin-driven approach for medical interventions.