Fares J. Abu-Dakka presents Geometry-aware representations for robot learning and control

On 2021-03-16 11:00:00 at https://feectu.zoom.us/j/99381567325
Robots are entering human environments such as houses, hospitals, and museums.
Such environments are highly unstructured, dynamic, and uncertain, making
explicit programming of required robot skills unfeasible. The difficulty is
particularly evident in manipulation tasks that require an encapsulation of its
characteristics in specific geometry constraints data type. In robotics, these
geometry constraints data types include: (i) Orientation data, encapsulated in
unit quaternions S3 or rotation matrices SO(3); (ii) Manipulability data,
encapsulated in Symmetric Positive Definite (SPD) matrices; and (iii) Impedance
data (stiffness and damping matrices), encapsulated in SPD matrices. Such data
types do not belong to the Euclidean space and thus the use of Euclidean space
arithmetic to operate over these data is inadequate. In this context,
differential geometry, or more specifically Lie group and Riemannian manifold
theories provide appropriate tools to learn such data, taking into account its
geometry constraints.


Fares J. Abu-Dakka is a candidate for the Robotics tenure-track position at the
Dept. of Cybernetics.
Responsible person: Petr Pošík