Klamt, Tobias: Planning Hybrid Driving-Stepping Locomotion for Ground Robots in Challenging Environments. - Bonn, 2020. - Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn.
Online-Ausgabe in bonndoc: https://nbn-resolving.org/urn:nbn:de:hbz:5-57320
@phdthesis{handle:20.500.11811/8276,
urn: https://nbn-resolving.org/urn:nbn:de:hbz:5-57320,
author = {{Tobias Klamt}},
title = {Planning Hybrid Driving-Stepping Locomotion for Ground Robots in Challenging Environments},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2020,
month = feb,

note = {Ground robots capable of navigating a wide range of terrains are needed in several domains such as disaster response or planetary exploration. Hybrid driving-stepping locomotion is promising since it combines the complementary strengths of the two locomotion modes. However, suitable platforms require complex kinematic capabilities which need to be considered in corresponding locomotion planning methods. High terrain complexities induce further challenges for the planning problem.
We present a search-based hybrid driving-stepping locomotion planning approach for robots which possess a quadrupedal base with legs ending in steerable wheels allowing for omnidirectional driving and stepping. Driving is preferred on sufficiently flat terrain while stepping is considered in the vicinity of obstacles. Steps are handled in a hierarchical manner: while only the connection between suitable footholds is considered during planning, those steps in the resulting path are expanded to detailed motion sequences considering the robot stability. To enable precise locomotion in challenging terrain, the planner takes the individual robot footprint into account. The method is evaluated in simulation and in real-world applications with the robots Momaro and Centauro. The results indicate that the planner provides bounded sub-optimal paths in feasible time. However, the required fine resolution and high-dimensional robot representation result in too large state spaces for more complex scenarios exceeding computation time and memory constraints.
To enable the planner to be applicable in those scenarios, the method is extended to incorporate three levels of representation. In the vicinity of the robot, the detailed representation is used to obtain reliable paths for the near future. With increasing distance from the robot, the resolution gets coarser and the degrees of freedom of the robot representation decrease. To compensate this loss of information, those representations are enriched with additional semantics increasing the scene understanding. We further present how the most abstract representation can be used to generate an informed heuristic. Evaluation shows that planning is accelerated by multiple orders of magnitude with comparable result quality. However, manually designing the additional representations and tuning the corresponding cost functions requires a high effort.
Therefore, we present a method to support the generation of an abstract representation through a convolutional neural network (CNN). While a low-dimensional, coarse robot representation and corresponding action set can be easily defined, a CNN is trained on artificially generated data to represent the abstract cost function. Subsequently, the abstract representation can be used to generate a similar informed heuristic, as described above. The CNN evaluation on multiple data sets indicates that the learned cost function generalizes well to realworld scenes and that the abstraction quality outperforms the manually tuned approach. Applied to hybrid driving-stepping locomotion planning, the heuristic achieves similar performance while design and tuning efforts are minimized.
Since a learning-based method turned out to be beneficial to support the search-based planner, we finally investigate if the whole planning problem can be solved by a learning-based approach. Value Iteration Networks (VINs) are known to show good generalizability and goal-directed behavior, while being limited to small state spaces. Inspired by the above-described results, we extend VINs to incorporate multiple levels of abstraction to represent larger planning problems with suitable state space sizes. Experiments in 2D grid worlds show that this extension enables VINs to solve significantly larger planning tasks. We further apply the method to omnidirectional driving of the Centauro robot in cluttered environments which indicates limitations but also emphasizes the future potential of learning-based planning methods.},

url = {https://hdl.handle.net/20.500.11811/8276}
}

Die folgenden Nutzungsbestimmungen sind mit dieser Ressource verbunden:

InCopyright