Application of Gaussian Processes to robot state estimation
Keywords
Loading...
Authors
Issue Date
2021-02-08
Language
en
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Lanillos & Cheng [8] introduced "a computational perceptual model based
on predictive processing that enables any multi sensory robot to learn, infer
and update its body confi guration when using arbitrary sensors with
Gaussian additive noise". Specifi cally, their algorithm works by first using
Gaussian Processes (GP) regression to learn the sensor generative model;
the mapping between the joint angles of the robot and the sensor values,
and then uses this learned generative model to generate a prediction about
the state of the robot. A well know problem with GPs is that they don't
scale well with a large number of data point or a high number of dimensions.
This may make this proposed approach difficult to use for complex robotic
systems. In order to asses whether this is a signifi cant concern and whether
it can be mitigated I examined how well a variety of their model (using an
Extended Kalman Filter) performs on a variety of different robotic systems
of increasing complexity. In addition I tested a sparse Gaussian Processes
model to investigate its potentia as a method to mitigate the poor scaling
of a traditional GP for complex robotic systems.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen