Regier, Peter: Robot navigation in cluttered environments. - Bonn, 2022. - Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn.
Online-Ausgabe in bonndoc: https://nbn-resolving.org/urn:nbn:de:hbz:5-65066
@phdthesis{handle:20.500.11811/9535,
urn: https://nbn-resolving.org/urn:nbn:de:hbz:5-65066,
author = {{Peter Regier}},
title = {Robot navigation in cluttered environments},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2022,
month = jan,

note = {Service robots are designed for non-industrial use to help people at home and in public spaces. Today, service robots perform a variety of tasks that range from distributing medical supply to clean the floor. The wide range of possible application and the complexity of service robots spark major research interest. The goal of the multidisciplinary research is to increase the autonomy of service robots. The capability to navigate in the environment is fundamental for service robots. In this thesis, we present new approaches for robot navigation in challenging indoor scenarios with little space to maneuver, many objects, and crowds of people. We collectively call such scenarios as ’cluttered’. The presented contributions increase the efficiency of robot navigation and allow for new capabilities of the robot to solve challenging problems. Initially, we introduce a method to incorporate clutter in to the navigation process. We extend the state of the art navigation cost function by considering the configuration and quantity of object in the vicinity of the robot. The result is a foresighted robot navigation behavior, that leads around the clutter when it is beneficial for the robot. The second approach predicts the time the robot needs to complete a navigation task, based on a 2D path. The estimation of time is an important feature for service robots to schedule their tasks, e.g., guiding groups in a museum. Unfortunately, due to the lack of a dynamic model the completion time is a priori unknown. Therefore we train a regression model that reliably predicts the completion time based on 2D path features. To achieve human-aware navigation through pedestrian crowds, we apply the social force model (SFM) to control the robot. Our new approach reduces the collision rate with pedestrians of the SFM controlled robot, while maintaining similar velocities. The method considers a set of motion commands and evaluates the outcome, by simulating the corresponding situation into the future. Since such a omniscient robot control is not feasible to use in a real world scenario, we train a network with the best evaluated control command from the simulation. Our method outperforms the standard control with the SFM and the network successfully mimics the improved omniscient behavior, but considers information that is available to the robot in a real world scenario. In the next approach, the robot learns a navigation policy, through reinforcement learning (RL). Self-learning approaches have the potential to reduce the amount of parameter tuning, that is required to operate a robot. The endeavor of tuning is time consuming and know-how-intensive. To reduce the workload in this context, we successfully apply RL for the task of learning a navigation policy from scratch. The learned policy is capable to reach the target location faster than the state of the art approaches, by optimizing the behavior when navigating close to obstacles. In some situations efficient and collision-free navigation is not enough to reach the goal, e.g., when the path is blocked by an object. To target those scenarios, our final approach combines object classification, fast 2D grid-based path planning, manipulation, and footstep planning to overcome objects by stepping over it or moving it to free the path to the goal. Our method is able to run on a small humanoid robot and finds paths through regions where traditional motion planning methods are not able to calculate a solution or require substantially more time. All navigation techniques presented in this thesis were thoroughly evaluated in various experiments. Our approaches advance the state of the art towards autonomous robot navigation in cluttered scenarios.},
url = {https://hdl.handle.net/20.500.11811/9535}
}

The following license files are associated with this item:

InCopyright