School of Science and Technology

[unavailable] Feature-based SLAM (multiple projects)

[Due to the current covid-19 situation this project is temporarily unavailable]

Introduction

These projects are all defined in collaboration with an industrial partnership with Electrolux. Note the disclaimer at the end regarding the availability of support from the company side.

Project 1: Simplify SLAM feature map for presentation

The thesis objective is to propose priors that would be used to complete and
enhance 3D features of the environment for improved environment presentation.

The Pure i9 robot uses its 3D laser sensor to detect flat vertical surfaces such
as walls and furniture sides. These are used in the SLAM system, as landmarks
for localisation. But they are not presented in the user maps, because they are
not always perfectly perpendicular (because of sensor noise), they appear to
have different heights (because the robot has a limited vertical field of view,
so the observed height of a wall will depend on the distance from which it is
observed) and they may be fragmented into multiple coplanar pieces.

We would like to develop and test experience-based priors to the SLAM landmarks,
covering mutual wall positioning, structural features, expected object height
and others. These landmarks will allow enhancing the map visualisation for the
end-user.

Project 2: Use of SLAM feature maps to recover from kidnapping

The thesis objective is to develop a method for robot recovery from kidnapping
using feature maps.

The Pure i9 robot uses its 3D laser sensor to detect flat vertical surfaces such
as walls and furniture sides. These are used in the SLAM system, as landmarks
for localisation. However, the system works under assumptions; there were no
excessive alterations of robots pose between the consecutive scans. Thus is the
robot will be kidnapped (picked up and moved to a different location), the robot
is not able to detect this and recover.

We would like to develop a system for recovering from robot kidnapping. The key
idea is to compare the configuration map landmarks with the configuration of
landmarks in the most recent scans.

Project 3: Monitor the EKF-SLAM to detect when the robot is lost

The thesis objective is to develop a method for detecting robot localisation
errors.

During normal operation, the robot sometimes re-observes SLAM features, and
sometimes observes new ones. But if the dead reckoning gets corrupted because of
slip (such as on certain directional carpets or if the robot gets stuck on a
threshold), the robot might fail to understand when it's re-observing features,
and instead add them as new ones. This can create a "parallel world" in the SLAM
map, and typically leads to problems when trying to navigate back to known
areas.

We would like to develop a system to monitor the SLAM map while the robot is
running and define robust criteria for detecting when this has happened, and
preferably also recover the correct pose of the robot.

COVID-19 disclaimer:

Because of Covid-19, the Electrolux office is expected to remain partially
closed during the spring, so we're proposing that the students are sitting in
Örebro. We would provide remote support, but it is, of course, possible to have
some meetings in person, either in Stockholm or in Örebro.

Because the projects will be done externally, we will not be able to share the
source code for the robots, but we can provide recorded data from runs with the
Pure i9 robot.

Contact