Event
Ph.D. Research Proposal Exam: Kasun Weerakoon Kulathun Mudiyanselage
Friday, December 8, 2023
10:00 a.m.
AVW 2224
Maria Hoo
301 405 3681
mch@umd.edu
ANNOUNCEMENT: Ph.D. Research Proposal Exam
Name: Kasun Weerakoon Kulathun Mudiyanselage
Committee:
Professor Dinesh Manocha (Chair)
Professor Pratap Tokekar
Professor Kaiqing Zhang
Date/time: Friday, December 8, 2023 10am-12pm
Location: AVW 2224
Committee:
Professor Dinesh Manocha (Chair)
Professor Pratap Tokekar
Professor Kaiqing Zhang
Date/time: Friday, December 8, 2023 10am-12pm
Location: AVW 2224
Title: Autonomous Mobile Robot Navigation in Complex Outdoor Environments
Abstract:
Autonomous mobile robots have increasingly been used for many real-world outdoor applications such as surveillance, search and rescue, planetary/space exploration, delivery, agricultural surveys, etc. However, complex and unstructured outdoor environments present a variety of terrain challenges for autonomous robot navigation such as unevenness, diverse surface conditions, and the existence of vegetation with different stiffness properties (e.g. tall grass, shrubs, trees). We investigate the aforementioned challenges from both robot perception and planning perspectives to introduce novel algorithms to autonomously navigate mobile robots in complex outdoor environments. First, we propose a Deep Reinforcement Learning (DRL) base algorithm that uses robot-centric elevation maps to navigate wheeled robots stably in highly uneven terrains. While developing such DRL algorithms, we study the sim-to-real transfer challenges of end-to-end RL approaches and possible strategies to minimize the reality gap. Then, we incorporate a Graph Neural Network (GNN) based proactive anomaly detection strategy to fuse multi-modal sensors such as camera and 3D LiDAR to perform reliability-aware navigation under challenging environment conditions such as occlusions, lighting changes, clutteredness, etc. Further, we introduce an Offline RL approach to identify terrain surface properties and vegetation density for legged robot navigation in unstructured outdoor environments. We deploy all our algorithms in real robot platforms to evaluate their performance in real-world outdoor environments
