AI-driven planning has greatly expanded the scope of autonomy, transforming how robots perceive, reason, and act in complex environments. Foundation models trained on Internet-scale multimodal data have been a key driver behind this boost in autonomy. However, deploying existing models and techniques in safety-critical applications remains challenging due to the lack of a structured methodology to enforce strict behavioral and safety requirements. These constraints are essential for reliable operation in dynamic, real-world settings.

Historically, formal specifications, particularly linear temporal logic (LTL) and signal temporal logic (STL), have been used to encode safety and mission rules for robotics. Although recent work has integrated such logics into learning-based frameworks, they are poorly aligned with modern pre-training-based paradigms that under-pin robot foundation models. Several theoretical and practical challenges must be overcome before these methods can be widely adopted.

My thesis work explores the integration of temporal logic specifications for modern AI-driven planning. It introduces methods to (1) decompose and efficiently evaluate complex STL specifications using modern auto-differentiation frameworks, (2) generalize specification design via Embedding Temporal Logic (ETL), which defines predicates over learned representations instead of explicit states, and (3) enforce STL specifications at inference time using SafeDec, a constrained decoding approach inspired by structured generation in large language models. Together, these contributions enable formal specification-driven planning in AI systems, ensuring safety in robotics and autonomous systems.

104 pages

Thesis Committee:
Eunsuk Kang (Chair)
Sebsatian Scherer
Changliu Liu
Karen Leung (University of Washington)

Nicolas Christin, Head, Software and Societal Systems Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu