Policy-Based Planning for Robust Robot Navigation

This thesis proposes techniques for constructing and implementing an extensible naviga-
tion framework suitable for operating alongside or in place of traditional navigation sys-
tems. Robot navigation is only possible when many subsystems work in tandem such as
localization and mapping, motion planning, control, and object tracking. Errors in any one
of these subsystems can result in the robot failing to accomplish its task, oftentimes requir-
ing human interventions that diminish the benefits theoretically provided by autonomous
robotic systems.

Our first contribution is DART, a method for generating human-followable navigation
instructions optimized for followability instead of traditional metrics such as path length.
We show how this strategy can be extended to robot navigation planning, allowing the robot
to compute the sequence of control policies and switching conditions which will maximize
the likelihood with which the robot will reach its goal. This technique allows a robot to
select a plan based on reliability in addition to efficiency, avoiding error-prone actions or
areas of the environment. We also show how DART can be used to build compact, topo-
logical maps of its environments, offering opportunities to scale to larger environments.

DART depends on the existence of a set of behaviors and switching conditions describ-
ing ways the robot can move through an environment. In the remainder of this thesis, we
present methods for learning these behaviors and conditions in indoor environments. To
support landmark-based navigation, we show how to train a CNN to distinguish between
semantically labeled 2D occupancy grids generated from LIDAR data. By providing the
robot the ability to recognize specific classes of places based on human labels, not only
do we support transitioning between control laws, but also provide hooks for human-aided
instruction and direction.

Additionally, we suggest a subset of behaviors that would provide DART with a suf-
ficient set of actions to navigate in most indoor environments and introduce a method to
learn these behaviors from teleloperated demonstrations. Our method learns a cost func-
tion suitable for integration into gradient-based control schemes. This enables the robot to
execute behaviors in the absence of global knowledge. We present results demonstrating
these behaviors working in several environments with varied structure, indicating that they
generalize well to new environments.