The ability to make compliant physical interactions enables robots to robustly contact with the environment. This plays a significant role in advancing robotic hands which must constantly interact with the world, often with large uncertainties and variation in its geometry, material, and constraints. In this thesis, I will introduce experimental methodologies on how compliant interactions can be designed, controlled, and assessed for robotic dexterous manipulation. Taking this experimental approach I have developed robotic systems that integrate compliance into its design and sense key interaction forces to quantitatively evaluate its behavior, which enabled the demonstration of three key results. Firstly, to explore assessment methodologies for compliant interactions, a sensorized raspberry physical twin is developed. The physical twin matches the mechanical properties of the real fruit and measures interaction forces during a harvest; enabling a direct comparison between a harvest performed by a robot and the human. By tuning the robot to match the human, the lab developed harvester is able to be directly transferred into the field with a success rate of 80%. In addition, the twin is also used as a benchmarking method to compare among robotic harvester designs. Secondly, the design of spatially distributed and tunable stiffness and its impact on manipulation tasks, is explored for an anthropomorphic hand. The ADAPT Hand incorporates compliant elements in the skin, fingers, and wrist, tuned to match human characteristics. A direct comparison with a fully rigid setup quantifiably shows that human-matched compliance improves robustness in uncertain environments, from single finger to full hand interactions. Using an open-loop motion grasp objects from a surface, a 93% success rate is achieved across 24 objects of varying sizes. Additionally, the hand displays self-organizing behavior, where interaction with objects induces different grasp types with 68% similarity to natural human grasps. Lastly, teleoperation of a compliant dexterous hand is presented to explore its capabilities for complex tasks. The human-matched kinematics and stiffness of the ADAPT Hand 2 enables contact-rich dexterous motions (e.g., multi-object grasping, light bulb screwing, in-hand cube rotation) to be intuitively performed using simple joint-to-joint mapping. Moreover, teleoperation is used for data collection to develop autonomous imitation learning based controllers for contact-rich tasks. Beyond these key results, the combination of the work performed in this thesis presents a holistic contribution to advancing robotic hands that leverage their compliance towards intelligent interactions.