Coverage and exploration planning with semantically informed maps

Correia Marques, Joao Marcos

Permalink

https://hdl.handle.net/2142/129831 Copy

Description

Title

Coverage and exploration planning with semantically informed maps

Author(s)

Correia Marques, Joao Marcos

Issue Date

2025-07-02

Director of Research (if dissertation) or Advisor (if thesis)

Hauser, Kris

Doctoral Committee Chair(s)

Hauser, Kris

Committee Member(s)

• Wang, Shenlong
• Hoiem, Derek
• Atanasov, Nikolay

Department of Study

Siebel School Comp & Data Sci

Discipline

Computer Science

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Metric-Semantic Reconstruction
• Semantic Mapping
• Coverage Planning
• Ultraviolet Disinfection
• Coverage Path Planning
• Manipulation-Enhanced Mapping
• Confidence Calibration
• 3D Mapping

Abstract

Outside of controlled factory environments, robots are expected to build their own environment representations to effectively plan and act in the world. This representation typically takes the shape of a semantic map constructed through some Simultaneous Localization and Mapping pipeline. In this thesis, I first present my work on a robotic application in the wild whose performance relies not just on the accuracy of the resulting metric-semantic world map, but also on its uncertainty: Semantically Informed Ultraviolet Disinfection Planning. Taking this application as an example of semantically informed coverage planning, I then present evidence that most of the sequential estimation strategies for semantic maps lead to overconfident label estimates, and that this overconfidence impacts the downstream performance of the robotic agents, both in disinfection and object goal navigation tasks. I also provide two novel confidence-preserving memory-efficient methods to perform online metric-semantic reconstruction in real time. I then extend the domain of coverage planning problems to include the coverage of interactive scenes by proposing a solution to the Manipulation-Enhanced Mapping problem, where a robot must efficiently survey an environment with a camera and manipulate objects in the environment to improve object visibility. This is achieved by leveraging neural networks to accelerate the belief updates of well-studied formalism for decision-making under uncertainty of Partially Observable Markov Decision Processes, enabling us to solve these problems in the belief space of metric-semantic maps. I finally conclude with a summary of the presented research, as well as promising future research directions in the areas of coverage planning, interactive scene representations and on further improving confidence calibration and uncertainty handling in 3D metric-semantic maps.

Graduation Semester

2025-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129831

Copyright and License Information

© 2025 Joao Marcos Correia Marques

Owning Collections