Graduate Thesis Or Dissertation
 

Learning-based Techniques for Fast and Robust Motion Planning

Pubblico Deposited

Contenuto scaricabile

Scarica il pdf
https://ir.library.oregonstate.edu/concern/graduate_thesis_or_dissertations/1z40m246h

Descriptions

Attribute NameValues
Creator
Abstract
  • Motion planning is a cornerstone of autonomous robots, enabling robots to safely and efficiently perform tasks such as package delivery, infrastructure inspection, and manipulation. However, as the field of robotics matures, robotic systems are being developed that (1) are challenging to analytically model, (2) require computationally expensive model-based controllers, and (3) have increasingly large numbers of degrees of freedom. Current planning approaches struggle to address these additional complexities during plan generation, resulting in unacceptably long planning times and plans that may be dynamically infeasible to execute. On the other hand, data-driven methods offer a promising way to utilize observed data to learn fast-to-compute models of complex systems, but they often lack the guarantees from more traditional methods. In this dissertation, we explore how learning-based methods can be used to augment classical motion planning algorithms to address these challenges. In our first contribution, we consider planning for a soft robotic arm that is difficult to analytically model. To address this deficiency, we develop a neural network that learns a residual between a simplified quasistatic model and observed data in order to account for unmodeled forces such as friction and load. We apply our approach to a pneumatically-driven soft robot manipulator and show that the resulting data-driven model reduces average end effector pose error compared to using the simplified model alone, while still being fast enough for motion planning. We incorporate this model into a RRT*-based planner and demonstrate that the plans generated using our model are more likely to be feasible when executed on hardware than plans generated using the simplified no load model. Next, we examine the challenge of planning for robots that necessitate computationally expensive model-based controllers to account for environments with nonlinear dynamics. We incorporate control into planning using a closed-loop RRT (CL-RRT*) algorithm. One significant bottleneck to CL-RRT* is that it must simulate the controller's response to various reference trajectories. To address this computational bottleneck, we develop a neural network that approximates a model predictive control (MPC) controller. This neural network controller is shown to substantially reduce computation time while outputting similar solution quality. This enables CL-RRT* to repeatedly query the controller every iteration without bottlenecking. Additionally, we extend CL-RRT* to leverage the parallelism offered by the neural network-based controller. We demonstrate parallelized CL-RRT* reduces planning time compared to its non-parallelized counterpart for an autonomous underwater vehicle. Finally, for our third contribution, we address the optimal motion planning for high degree of freedom systems such as mobile manipulators. We develop the Warm-starting Adaptively Informed Trees (WAIT*) algorithm, which utilizes an ensemble of sub-optimal and learning-based planners in combination with informed search. While these fast but non-optimal planners may yield high cost solutions with potentially invalid state transitions, the states that these paths are comprised are more likely to be part of a solution than uniform sampling. The informed planner uses these salient states to quickly find an initial solution, which is used to eliminate regions of the search space where a lower cost solution cannot exist. Therefore, by finding an initial solution faster using salient states, WAIT* is able to spend more time sampling in an informed region composed only of states that can yield a lower cost solution and thus converge more quickly towards the optimal solution. Our approach improves planning success rate compared to a standard informed tree planner, while still having the ability to iteratively improve solution quality after finding the initial solution.
License
Resource Type
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Advisor
Committee Member
Dichiarazione dei diritti
Funding Statement (additional comments about funding)
  • This research was funded in part by the National Science Foundation (NSF award IIS-1734627) and the Office of Naval Research (ONR award N0014-21-1-2052 and ONR/NAVSEA contract N00024-10-D-6318/DO#N0002420F8705).
Publisher
Peer Reviewed
Language
File Format

Le relazioni

Parents:

This work has no parents.

In Collection:

Elementi

Thumbnail Titolo Data caricata Visibilità Azioni
file details ChowScott2024.pdf 2024-03-12 Pubblico Scaricare