Learning for open-world mobile robots

Deploying robots from a lab setting to open-world environments requires understanding semantic information and leveraging large data sources. Various paradigms have been introduced to tackle semantic visual-goal navigation, where an agent is placed in a random environment and must reach a goal. We first decompose this task into two components: (1) a data-driven exploration policy that learns semantics and environmental relations and (2) a geometric-based policy specialized for goal-directed navigation. Beyond this decomposition, we investigate whether further structure can enhance performance. To this end, we retrain the exploration policy with guidance from the geometric policy. Additionally, we explore a sim-to-real approach to improve state estimation for legged robots, enabling robust odometry prediction across diverse scenarios. This thesis presents real-world experiments supporting each of these works in mobile robotics. Future work is finally discussed towards the development of a foundation model for navigation.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Wasserman, Justin

Contributors dc:contributor

• Chowdhary, Girish
• Driggs-Campbell, Katie
• Schwing, Alexander
• Wang, Shenlong

• robotics
• embodied ai
• simulator
• navigation
• open-world
• artificial intelligence
• computer vision
• machine learning

Statement dc:rights

• Copyright 2025 Justin Wasserman

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129393

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129393