Collaborative Modeling and Repair of Reactive Model-Based Plans with Natural Language
Name
zagara-czagara-sm-aeroastro-2026-thesis.pdf
Size
3.9 MB
Format
Adobe PDF
Checksum (MD5)
5ab25f21b7b460e85303490bc4aeb481
Author(s)
Zagara, Cecilia Elizabeth
Advisor(s)
Williams, Brian C.
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
As robotics becomes more accessible to the scientific community, the algorithms driving it must follow suit. In high-stakes domains such as underwater exploration, accurate activity planning is mission-critical; in unforgiving environments where errors are costly, system guarantees are a prerequisite for safety. While the automated planning community offers sophisticated algorithms capable of managing uncertainty, concurrency, and risk constraints, these tools require problems to be defined in rigid formal languages. Such languages are notoriously time-consuming and errorprone for experts and novices alike. Conversely, Large Language Models (LLMs) provide an intuitive natural language interface and vast common-sense knowledge for drawing conclusions; however, they lack the formal rigor required for high-stakes tasks. LLMs are hindered by hallucinations and a limited capacity for complex temporal reasoning, making them unable to provide the completeness or correctness guarantees essential for long-horizon planning. This thesis introduces XR-Uhura, a collaborative framework that synthesizes correct, executable plans from natural language state plan descriptions by integrating the strengths of users, automated planning algorithms, and LLMs. The framework resolves three critical challenges in interactive planning. First, it facilitates collaborative modeling and candidate plan construction, in which the user provides real-world grounding, and the LLM iteratively identifies missing information required for planning. Second, it employs formal verification, utilizing automated planners to detect conflicts and ensure plan correctness. Finally, the framework enables interactive plan repair: the LLM translates formal conflicts into natural language, allowing users and the LLM to diagnose and resolve them.
MIT Department
Massachusetts Institute of Technology. Department of Aeronautics and Astronautics
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
