The level of refined operation of a robotic hand depends on its fingertip tactile perception performance. To enhance the tactile perception performance of FBG-based robotic fingertips, a diagonal cross-shaped FBG tactile perception unit was developed, featuring a flexible packaging structure composed of a graphene-silicone composite material. This design addressed key challenges such as poor thermal conductivity and the cross-sensitivity between contact force and contact temperature commonly observed in conventional FBG soft perception units. To resolve the coupling issue between force and temperature, a decoupling method based on an Osprey optimization algorithm optimized convolutional neural network (OOA-CNN) was proposed. First, simulation analyses were conducted to compare the temperature response and strain response of FBG sensors embedded in graphene-silica gel composite packaging versus pure silica gel packaging. Then, experimental analyses were performed to investigate the impact of different graphene mass fractions (1%, 1.5%, 2%, 2.5%, and 3%) on the thermal conductivity of composite materials. Using a three-fingered robotic hand, fingertip perception experiments were carried out to evaluate the sensitivity of the FBG tactile perception unit to both contact force and contact temperature. Finally, coupled analysis was performed on the composite perception data of contact force and temperature. The decoupling performance of the proposed OOA-CNN model was validated through comparative experiments against a standard CNN model and a least squares method. Simulation and experimental results show that incorporating 1.5% mass fraction graphene as a thermally conductive filler in the silicone matrix significantly enhances its thermal conductivity while preserving the FBG′s tactile sensing performance. The FBG tactile unit exhibited a sensitivity of 31.281 pm/N to contact force and 10.787 pm/℃ to contact temperature. Furthermore, the OOA-CNN decoupling model has a better decoupling effect compared to the least squares method and the CNN decoupling model, with an average absolute error reduction of 40.3% for contact temperature and 41.33% for contact force.