As technology continues to evolve, robots are becoming an intrinsic part of our lives. While robots surpass human capabilities in precision and speed, they are far from matching humans' ability to adapt to unexpected changes. Humans can quickly and compliantly respond to uncertainties and perform complex tasks such as dexterous manipulation. Robots in human-centric environments need to be equipped with human-like capabilities, primarily when such a machine interacts with us or the objects around us. Therefore, the thesis aimed at (i) obtaining a precise controller with human-like compliant capabilities and, with such a controller, (ii) designing a control pipeline to perform dexterous manipulation.

Learning, assessing, and constantly updating the underlying robot's dynamics is the first step to obtaining a precise and compliant controller. Accordingly, we present methods to learn and update dynamics models in the first two parts of the thesis. The first part of the thesis investigates supervised machine learning techniques for learning the inverse dynamic model of a robot prior to task execution. We propose a method for incrementally exploring a robot's configuration space and maximizing the information of the collected data.

In the second part, we focus on assessing and updating dynamic models during and after task execution. We offer a method that augments episodic model updates with online adaptation. We propose combining traditional model-based controllers with a learned residual inverse dynamics model. Then, we introduce an adaptive control law that adjusts the control reference online to account for model uncertainties and unforeseen disturbances.

We dedicate the third part of the thesis to developing a compliant control pipeline to achieve human-like dexterity tasks. We introduce a new robust and synchronized planning schematic for grasping and manipulating tasks. Our approach is to control and synchronize fingers based on dynamical systems. We combine our adaptive controller with joints' impedance regulation to guarantee high tracking accuracy and adapt to dynamic changes. We showcase that in conjunction with learning from human demonstration, our controller provides a robust solution for more complicated manipulations such as finger gaiting.

Lastly, we use the developed robotic hand controller in two applications in the human-centric environment. The task in the first application is to increase the dexterity of robotic prosthetic hands (RPHs) for individuals with a hand amputation using a new teleoperated-control scheme via electromyography. In the second application, we perform tactile surface exploration of apriori unknown objects through a novel informative path planning exploration strategy.

In summary, this thesis offers a robot controller, compliant in interaction and faithful in tracking, for dexterous grasp and manipulation tasks. Results indicate substantial accuracy improvement over traditional approaches when using our incremental method for configuration exploration. We confirmed by various experiments that our residual model learning procedure could learn unmodeled dynamics in a hand full of trials and adapt online to perturbations in unpredictable environments. Finally, with extensive experiments and two applications, we showed that introducing a coordinated multi-finger system to our controller provides a robust solution for grasping and manipulating problems in uncertain environments.