PhD Proposal by Florian M. Hauer
Ph.D. Thesis Proposal
Florian M. Hauer
Advisor: Prof. Panagiotis Tsiotras
2 pm, Wednesday, June 29th, 2016
Montgomery Knight 325
Multi-Resolution and Machine Learning Techniques to Improve Path-Planning Algorithms
Path-Planning algorithms are used to find a feasible trajectory for systems from a starting configuration to a goal configuration while avoiding obstacles, and often minimizing a cost function. They are widely studied, as they constitute a major component of most robotic and autonomous systems. However, such algorithms do not scale well as the system complexity or the dimension of the search space increases, and represent the computational bottleneck of a typical autonomous system.
This research proposes different techniques to reduce the performance slowdown in high dimensions. For discrete path-planning, we propose a multi-scale decomposition that can provide information for large regions of the space without having to explore them at the finest resolution. Such a decomposition is well-suited to guide a planner through relevant regions of the search space. For continuous search spaces, randomized algorithms, such as rapidly exploring random trees, are often used and rely on sampling “good” points in the search space. Using optimization techniques, we show how samples can be slightly adjusted to become more relevant, maximizing the information gathered by the planner with a fixed number of samples, for example. We also use machine learning techniques to estimate computationally expensive results required for the execution of these algorithms, such as the solution of an optimal control problem. The planning problem is thus solved much faster, at the cost of an approximated solution. Finally, an experimental platform will allow us to compare the different techniques on a real-world system.
Prof. Panagiotis Tsiotras, School of Aerospace Engineering
Prof. Eric Feron, School of Aerospace Engineering
Prof. Evangelos Theodorou, School of Aerospace Engineering