Projects

Phase 2 Projects

Active exploration in the high-dimensional data of an artificial skin


Project leaders: Patrick van der Smagt, (Oberpfaffenhofen), Alex Graves (München)

Researchers: Christian Osendorfer (München), Sebastian Urban (München)

Administration:

Associates:

Summary:

We target the separation of nonlinearly interacting causes in data through learning using autonomous active exploration. We can measure data s from a function f : X -> S without having access to x. We can furthermore measure data from a function g : X × O -> S, where we know that for one specific Ø ? O, g(x, Ø) = f (x) holds for all x ? X. From measured data describing the function g, how can we infer x and o?
Apart from being theoretically appealing, this topic has many practical applications. We will focus on the following real-world application: a skin-sensorised robotic gripper. With x being finger positions and o the properties of a hand-held object, let f (x) be the readout of the tactile skin due to finger egomotion—in our setup, the skin also spans the creases of the fingers. Let g(x,o) be the readout from the tactile skin when a particular object is hand-held; g and f are identical when there is no object (represented as Ø above) in the hand. We are interested in obtaining both the finger position x and object properties o given data from g.
We propose to decause g by autonomous active exploration in the skin data S. By moving the fingers, we can actively search for interesting areas in the codomain of g. In particular, we are looking for those s ? S for which our finger position and object information cannot yet be separated well enough. We do this by combining unsupervised learning methods with temporal learning in a reinforcement learning framework in which the system optimises its knowledge of X and O by exploring S.

Phase 1 Projects

Active exploration in the high-dimensional data of an artificial skin


Project leaders: Patrick van der Smagt, (Oberpfaffenhofen), Alex Graves (München)

Researchers: Christian Osendorfer (München), Sebastian Urban (München)

Administration:

Associates:

Summary:

We target the separation of nonlinearly interacting causes in data through learning using autonomous active exploration. We can measure data s from a function f : X -> S without having access to x. We can furthermore measure data from a function g : X × O -> S, where we know that for one specific Ø ? O, g(x, Ø) = f (x) holds for all x ? X. From measured data describing the function g, how can we infer x and o?
Apart from being theoretically appealing, this topic has many practical applications. We will focus on the following real-world application: a skin-sensorised robotic gripper. With x being finger positions and o the properties of a hand-held object, let f (x) be the readout of the tactile skin due to finger egomotion—in our setup, the skin also spans the creases of the fingers. Let g(x,o) be the readout from the tactile skin when a particular object is hand-held; g and f are identical when there is no object (represented as Ø above) in the hand. We are interested in obtaining both the finger position x and object properties o given data from g.
We propose to decause g by autonomous active exploration in the skin data S. By moving the fingers, we can actively search for interesting areas in the codomain of g. In particular, we are looking for those s ? S for which our finger position and object information cannot yet be separated well enough. We do this by combining unsupervised learning methods with temporal learning in a reinforcement learning framework in which the system optimises its knowledge of X and O by exploring S.