Felix Endres 论文下载
Technische Fakult¨ at
Albert-Ludwigs-Universit¨ at Freiburg
Betreuer: Prof. Dr. Wolfram Burgard
简介：In recent years, commercially available mobile robots, that operate in indoor environments, have found their ways into private homes, ofﬁce environments, and industrial settings.They fulﬁll important duties such as transportation, telepresence, and cleaning.While some tasks, such as vacuum cleaning, can be achieved with rudimentary perception, for more complex tasks sophisticated perception capabilities are fundamental for autonomous robots.
Even for the mentioned vacuum cleaning, improved perception al-
lows for substantial increases in efﬁciency. In this thesis, we investigate novel approaches
for the modelling of indoor environments for robot navigation. Being an important foun-
dation for higher level skills, a particular focus lies on simultaneous localization and
mapping (SLAM), which allows a robot to construct a model of its environment dur-
ing operation. In the context of SLAM, we develop an approach for RGB-D cameras,
that captures dense 3D maps for robot navigation. For this SLAM system, we propose
novel methods to increase the accuracy of the trajectory estimation and the robustness
against misassociations during individual motion estimates. Further, we address a major
limitation on the hardware side of RGB-D cameras, namely the limited ﬁeld of view. We
investigate SLAMwith multiple RGB-D cameras and develop an approach for automated
extrinsic calibration of RGB-D cameras via SLAM. We further propose an extension of
RGB-D sensors with mirrors to bisect the ﬁeld of view into two roughly opposite views.
While this does not increase the overall information perceived, we show that the divided
ﬁeld of view is beneﬁcial in the context of SLAM. Additionally, we exploit the structural
properties of this catadioptric extension to constrain the mentioned calibration method,
such that planar motion of the robot is sufﬁcient for online calibration of the two views.
To autonomously access all areas in a private home or ofﬁce, a further key skill for robot
navigation is the operation of doors. In this context, we extend the state of the art by novel
methods for learning a model of the kinematics and dynamics of a door. We demonstrate
that the knowledge about the dynamics of a door allows the robot to accurately predict
the motion of the door from inertia. We use this ability to employ a door opening strategy
with low requirements on the dexterous workspace of the manipulator. To show the bene-
ﬁts of the approaches proposed in this thesis, we thoroughly evaluate them in experiments
with real robots and real sensor data. Overall, the proposed approaches lower the cost of
the sensing equipment and the required complexity of the manipulator. These factors are
particular important for commercial robots targeted at households and small businesses.