This study evaluates the effectiveness of a deep learning-based grasp detection model, traditionally used with parallel plate grippers, when applied to a dexterous soft anthropomorphic robotic hand. By integrating the 6IMPOSE deep learning framework, this research leverages the soft hand’s intrinsic compliance and larger contact surface to enhance grasp performance in complex and dynamic environments. A novel geometric mapping algorithm was developed to discern between pinch and power grasps, adjusting the end-effector’s geometry dynamically during the grasping process. Comparative experimental evaluations were conducted to assess the performance of this approach against traditional gripping mechanisms. Results indicate that the adapted models improve grasp stability and adaptability in scenarios characterized by clutter and variability. The contributions of this work include the algorithm for mapping traditional gripper grasps to anthropomorphic configurations and comprehensive performance analysis.