Title: Informed Exploration Algorithms For Robot Motion Planning And Learning

Sagar Suhas Joshi

Robotics Ph.D. Candidate

Georgia Institute of Technology

Email: sagarsjoshi94@gatech.edu

Date: Monday, April 11, 2022

Time: 11:00 AM to 1:00 PM (ET)

Meeting Link: https://gatech.zoom.us/j/94382308397

Committee:

Dr. Panagiotis Tsiotras (Advisor) -- School of Aerospace Engineering, Georgia Tech

Dr. Seth Hutchinson -- School of Interactive Computing, Georgia Tech

Dr. Mathew Gombolay --School of Interactive Computing, Georgia Tech

Dr. Le Song-- Mohamed Bin Zayed University of Artificial Intelligence, UAE

Dr. Byron Boots -- School of Computer Science and Engineering, University of Washington

Abstract:

Sampling-based methods have emerged as a promising technique for solving robot motion-planning problems. These algorithms avoid a priori discretization of the search-space by generating random samples and building a graph online. While the recent advances in this area endow these randomized planners with asymptotic optimality, their slow convergence rate still remains a challenge. 

One of the reasons for this poor performance can be traced to the widely used uniform sampling strategy that naively explores the entire search-space. Having access to an intelligent exploration strategy that can focus search, would alleviate one of the critical bottlenecks in speeding up these algorithms. This thesis endeavors to tackle this problem by presenting exploration algorithms that leverage different sources of information available during planning time.

First, a non-parametric sampling technique is proposed for the basic case of geometric path-planning in uniform cost-spaces. This method employs sparsification and utilizes heuristics and prior obstacle data to conduct a prioritized search.

Second, the uniform cost-space assumption is relaxed, and a “Relevant Region” sampler is proposed for efficient exploration on general cost-maps. The proposed framework reduces dependence on heuristics by utilizing the planner’s tree structure information.

Next, the Relevant Region framework is extended by proposing a locally exploitative procedure that formulates sampling as an optimization problem. This technique generates new samples to improve the cost-to-come value of vertices in a neighborhood.

Fourth, the geometric setting is relaxed and the problem of efficient exploration for the case of kino-dynamic motion planning is discussed. A “Time-Informed” exploration procedure that focuses search by leveraging ideas from reachability analysis, while still maintaining asymptotic optimality guarantees, is presented.

Fifth, this thesis also explores the problem of intelligent query sampling for data generation to improve the performance of learning-based planners.

Finally, this thesis develops a software stack for implementation of the proposed planning algorithms on a 1/10 th scale RACECAR robot.