Robotic grasping of objects semi-buried in granular media (GM) is challenging due to particle resistance. Direct grasping requires excessive contact force to generate sufficient friction, easily damaging the object. In this work, we first design an adaptive gripper with multimodal sensors for excavation, integrating a binocular camera, tactile sensor array (TSA), and inertial measurement unit (IMU) to detect the object contour, the contact state, and fingertip motion. Based on the gripper, we develop a teleoperation system with visual-tactile feedback to collect high-quality human demonstration data. We then propose a visual-tactile fusion-driven diffusion policy to learn efficient excavation actions. Experiments including robotic end-pose tracking and object excavation are conducted to validate the performance of the system and the policy.