Motion pattern prediction in dynamic environments

Traditionally, motion planning algorithms have tackled the challenge of navigation in dynamic environments by approximating a robot's configuration space through a graph representation. This involves predicting or computing the trajectories of obstacles and finding feasible paths via a pathfinding algorithm. In our work, we strive to enhance the performance of these subproblems by learning to identify regions critical to dynamic environment navigation. We present a novel methodology for constructing sparse probabilistic roadmaps, a two stage approach that combines a self-supervised learning method for recognizing the topology and geometries indicative of motion patterns in dynamic settings, and a sampling-based strategy for leveraging these learned features. The result is the creation of neural networks capable of predicting the probability of occupancy of a given region and Avoidance Critical Probabilistic Roadmaps (ACPRMs), which leverage these insights to significantly improve navigation performance. ACPRMs have shown remarkable performance, demonstrating up to five orders of magnitude improvement over grid-sampling in multi-agent scenarios and surpassing competitive baselines by up to ten orders of magnitude in multi-query situations.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Arias, Felipe Felix

Contributors dc:contributor

• Amato, Nancy M

• Robotics
• Navigation
• Motion Planning
• Dynamic Environments
• Multi-agent Systems
• Machine Learning

Statement dc:rights

• Copyright 2023 Felipe Arias

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/122055

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F122055