Traversability prediction and navigation for unstructured environments

Accurate and robust navigation in unstructured environments is a critical challenge for autonomous robots. Currently, methods rely on predetermined paths, heuristics-based obstacle avoidance, or supervised learning by handcrafted labels. With the goal of improving navigation and control of autonomous agents in unstructured environments, this dissertation presents a series of works that leverage sensor fusion, machine learning, and predictive modeling techniques to provide better decision making and accurate control for robotic systems. The proposed methods enable robots to navigate semi-structured and unstructured environments efficiently while avoiding navigational failures. By learning traversability parameters and incorporating them into navigation systems, we demonstrate significant improvements in autonomous performance. The first contribution, CropNav, addresses the challenges of navigating under agricultural canopies where GNSS signals are unreliable. By autonomously switching between sensing modalities such as LiDAR-based row-following and waypoint tracking, CropNav enables seamless navigation inside and outside crop rows. Additionally, an online optimization approach is introduced to learn traction coefficients and recover from navigation failures. This system extends autonomous navigation time with reduced human intervention, achieving up to 750~m per intervention compared to GNSS-based navigation. Subsequent works explore the integration of deep learning with model predictive control (MPC) for enhanced performance while preserving stability. LBMPC introduces a dual-timescale adaptation mechanism for real-time uncertainty estimation using deep neural networks. WayFAST and WayFASTER build on these ideas by employing self-supervised learning to predict traversable paths in unstructured environments using RGB and depth data. These methods demonstrate improved data efficiency and the ability to handle challenging terrains such as snow or tall grass. Finally, this dissertation introduces TRAIL and ZEST, two novel approaches addressing limitations in generalization and adaptability. TRAIL learns traversability representations invariant to robot embodiment, enabling deployment across diverse platforms with improved safety through uncertainty prediction. ZEST leverages large multimodal large language models (LLMs) for zero-shot traversability prediction, generating global traversability maps directly from sensory inputs. Together, these contributions advance the state-of-the-art in autonomous navigation by enabling robust and risk-aware systems capable of operating in diverse and challenging environments.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Valverde Gasparino, Mateus

Contributors dc:contributor

• Chowdhary, Girish
• Hoiem, Derek
• Driggs-Campbell, Katherine
• Becker, Marcelo

• Autonomous Navigation
• Traversability
• Deep Learning
• Perception
• Robotics

Statement dc:rights

• Copyright 2025 Mateus Valverde Gasparino

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129688

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129688