• human grasping
• autonomous manipulation
• robot grasping

Envisioning in the (not-so-far) future robots which are able to autonomously grasp and manipulate objects, interacting with humans and their environment, is becoming more and more concrete as the research in the field brings new, promising results.
A path started more than 50 years ago, with direct inspection of humans performing various sort of manipulative tasks, passing through categorization (resulting in the so called "grasp taxonomies"), and then towards building an hardware as close as possible to human hands in order to be able to mimic their behavior.
Grasp modeling was used to this end, but the initial hypothesis needed for simplifying such a complex problem has been to have isolated contact points, happening only between the distal phalanxes of the robot hand and an external object.
Recently, the concepts of soft interaction and soft robotic hands is becoming increasingly widespread: they started to change the paradigm, from timid, contact-based interactions with the objects to be manipulated, to daring, intense whole-hand interactions also involving the surrounding environment.
Disparate grasp planning algorithms have been developed mainly for the former kind of hands, although some attempts have begun to sprout also for the latter ones.
Moreover, increasing the autonomy of the system, allowing high-level specifications to be interpreted and executed, has been studied as a necessary step towards simpler, more natural task definition along the path on the way to interaction of the robot with a human-centered world.
This thesis presents some advancements in aforementioned building blocks, towards the increase of robotic manipulation ability.
The first part includes a novel, parametric kinematic model which can be adapted to different subjects and takes into account the relative motion between skin and bones in order to accurately reconstruct the hand motion of a human performing grasping and manipulation tasks.
Data from various subjects have been collected and analyzed, and a new clustering algorithm is used to obtain a data-driven grasp taxonomy.
In the second part, the problem of how to transfer the knowledge gathered from humans to robotic systems is faced, and two different ways are explored: recording humans performing grasping motions while "wearing" the robotic end-effector as a tool, and a learning algorithm capable of working with soft robotic hands which generalizes successful example grasps to new scenarios.
Finally, the third part involves giving the robot an increased autonomy, using an abstraction layer which makes robot end-effectors and fixed environment elements alike, each with its own interaction primitives to act on the object; it is shown how it is possible to translate higher level instructions into a sequence of low level actions, which can be then executed by the robotic system.