Reproduce “Automated Generation of MDPs Using Logic Programming and LLMs for Robotic Applications” on OpenResearch

About this paper

Enrico Saccon, Davide De Martini, Matteo Saveriano, Edoardo Lamon, Luigi Palopoli, Marco Roveri

A framework autonomously creates probabilistic robotic policies from natural language descriptions by leveraging large language models to construct a Prolog knowledge base, from which a Markov Decision Process is built and optimal policies are synthesized via model checking for real-world deployment.

Key insights

• Large language models can effectively bridge the gap between informal natural language and formal logic programming (Prolog) for robotic planning, inferring complex elements like nuanced reward structures.
• Integrating LLMs with formal methods, specifically Prolog KBs and model checking, provides interpretability and ensures the generation of optimal, verifiable probabilistic policies for uncertain environments.
• The framework enables flexible policy generation tailored to diverse optimization objectives, such as maximizing success probability or minimizing specific rewards, allowing for task-specific trade-offs in robotic behavior.

Summary by alphaXiv.