TU Berlin

Robotics and Biology LaboratoryLearning Switching Kalman Filters (SKF)

Fachgebietslogo RBO

Page Content

to Navigation

Learning Switching Kalman Filters (SKF)

Motivation

Robots need to solve different sub-problems on their way to solve complex manipulation tasks. As robots progress through these consecutive stages of the overall task, they need to understand which of the sub-problems they need to solve at the current point in time.

Each of these sub-problems is often characterized by unique sensorimotor dynamics that are different from the dynamics of the other tasks. If we can identify the sensorimotor dynamics at the current point in time, then we can also identify what sub-problem needs to be solved.

Description of Work

A tool that can be used to identify the dynamics that currently govern the system, given a candidate set of dynamics, are Switching Kalman Filters (SKF). SKF can be seen as mixture of Kalman Filters (KF), or as a combination of Hidden Markov Models (HMM) and KF. Although there has been research on how to do inference in SKF, it is not yet clear what the best way is to learn these models. In this thesis you will learn about and develop new approaches to learn SKF models. You will evaluate these models on real world manipulation data.

Contact: Manuel Baum, Oliver Brock

Navigation

Quick Access

Schnellnavigation zur Seite über Nummerneingabe