Data-driven Learning of the Manifolds of a Robotic Manipulator

Recently, many hand-like robotic manipulators with high degrees of freedom have been developed. Specifically, anthropomorphic robotic hand models such as the Shadow Hand by Shadow Robot are highly redundant and offer significant potential for effective object manipulation. However, controlling such systems is difficult due to the high dimensionality of the joint space. Paradoxically, those kinds of manipulations are deemed easy for humans who do not control each joint of their hand individually but rather use a low-dimensional representation of their hand to perform the manipulation, often referred to as synergies. In this work, we propose learning the manifolds in the joint space resulting from synergies via a robotic manipulator using an autoencoder. We will see how the learned manifolds can be analyzed and used to simplify the control of the manipulator.